Dissertations / Theses on the topic 'Radiomics analysis'

Radiomics, a non-invasive and quantitative mining medical imaging information method, could extract molecular biological features and enormous feature combinations to customise individualised treatment and solve the problem of heterogeneity, satisfying the standards of precision medicine. However, it faces many challenges in the feature selection process, including redundant features, irrelevant features and the overfitting risk. More important, people know little about radiomics biological background and its connection to radiology, so it is difficult to apply radiology directly to medicine as it lacks interpretability. The core of this thesis is radiomic biology analysis that connects radiomic imaging information with molecular biology information to achieve a medical “gold standard” for cancer management. Developing methods to succeed in the feature selection process of data of varying dimensions is the main goal of this paper. Our major contributions in this thesis can be summarised as below: 1. We firstly proposed an unsupervised learning framework to guide supervised learning in the reduction of feature dimensions from large cohorts Non-Small Cell Lung Cancer data (NSCLC) on both clinical data and radiomic data for survival prediction. 2. An interpretable machine learning approach measures the contribution of features for each case and the connection of radiomics to its underlying biological features to make clinical decisions in leukemia and breast cancer cases. The weight of the feature can be estimated by measuring the distance of the approximate perturbation centre. 3. Based on the framework of feature selection that we proposed, to ensure the fairness and stability of the data split when processing classification results, cross-validation is embedded in the training process. We further propose a traversal selection method, optimising the computational complexity of the selection process to obtain the most robust feature set.

[ES] En los últimos años, los investigadores han intentado explotar la información de las imágenes médicas a través de la evaluación de parámetros cuantitativos para ayudar a los clínicos con el diagnóstico de enfermedades. Esta práctica ha sido bautizada como radiomics. El análisis de texturas proporciona una gran variedad de parámetros que permiten cuantificar la heterogeneidad característica de diferentes tejidos, especialmente cuando se obtienen de imagen por resonancia magnética (IRM). Basándonos en esto, decidimos estudiar las posibilidades de los parámetros texturales extraídos de IRM para caracterizar varios trastornos cerebrales. El potencial de las texturas se analizó con enfoques de aprendizaje automático, usando diferentes clasificadores y métodos de selección de características para hallar el modelo óptimo para cada tarea específica. En esta tesis, la metodología radiomics se usó para realizar cuatro proyectos independientes. En el primer proyecto, estudiamos la diferenciación entre glioblastomas multiformes (GBMs) y metástasis cerebrales (MCs) en IRM convencional. Estos tipos de tumores cerebrales pueden confundirse al diagnosticarse, ya que presentan un perfil radiológico similar y los datos clínicos pueden no ser concluyentes. Con el fin de evitar procedimientos exhaustivos e invasivos, estudiamos el poder discriminatorio de texturas 2D extraídas de imágenes de referencia T1 filtradas y sin filtrar. Los resultados sugieren que los parámetros texturales proporcionan información sobre la heterogeneidad de los GBMs y las MCs que puede servir para distinguir con precisión ambas lesiones cuando se utiliza un enfoque de aprendizaje automático adecuado. En el segundo proyecto, analizamos la clasificación de las MCs según su origen primario en IRM de referencia. En un porcentaje de pacientes, las MCs son diagnosticadas como la primera manifestación de un tumor primario desconocido. Con el fin de detectar el tumor primario de una forma no invasiva y más rápida, examinamos la capacidad del análisis de texturas 2D y 3D para diferenciar las MCs derivadas de los tumores primarios más propensos a metastatizar (cáncer de pulmón, cáncer de mama y melanoma) en imágenes T1. Los resultados mostraron que se logra una alta precisión al usar un conjunto reducido de texturas 3D para diferenciar MCs de cáncer de pulmón de MCs de cáncer de mama y melanoma. En el tercer proyecto, evaluamos las propiedades del hipocampo en IRM para identificar las diferentes etapas de la enfermedad de Alzheimer (EA). Los criterios actuales para diagnosticar la EA requieren la presencia de déficits cognitivos severos. Con la idea de establecer nuevos biomarcadores para detectar la EA en sus primeras etapas, evaluamos un conjunto de texturas 2D y 3D extraídas de IRM del hipocampo de pacientes con EA avanzada, deterioro cognitivo leve y normalidad cognitiva. Muchos parámetros de textura 3D resultaron ser estadísticamente significativos para diferenciar entre pacientes con EA y sujetos de las otras dos poblaciones. Al combinar estos parámetros con técnicas de aprendizaje automático, se obtuvo una alta precisión. En el cuarto proyecto, intentamos caracterizar los patrones de heterogeneidad del ictus cerebral isquémico en IRM estructural. En IRM cerebral de individuos de edad avanzada, algunos procesos patológicos presentan características similares, como las lesiones por ictus y las hiperintensidades de la sustancia blanca (HSBs). Dado que los ictus afectan también al tejido adyacente, decidimos estudiar la viabilidad de texturas 3D extraídas de las HSBs, la sustancia blanca no afectada y las estructuras subcorticales para diferenciar sujetos afectados por ictus lacunares o corticales visibles en IRM convencional (imágenes T1, T2 y FLAIR) de sujetos sin ictus. Las texturas no sirvieron para diferenciar ictus corticales y lacunares, pero se lograron resultados prometedores para discernir pacientes qu
[CAT] En els últims anys, els investigadors han intentat explotar la informació de les imatges mèdiques a través de l'avaluació de nombrosos paràmetres quantitatius per ajudar els clínics amb el diagnòstic i la valoració de malalties. Aquesta pràctica ha sigut batejada com radiomics,. L'anàlisi de textures proporciona una gran varietat de paràmetres que permeten quantificar l'heterogeneïtat característica de diferents teixits, especialment quan s'obtenen a partir d'imatge per ressonància magnètica (IRM). Basant-nos en aquests fets, vam decidir estudiar les possibilitats dels paràmetres texturals extrets d'IRM per caracteritzar diversos trastorns cerebrals. El potencial de les textures es va analitzar amb mètodes d'aprenentatge automàtic, usant diferents classificadors i mètodes de selecció de característiques per trobar el model òptim per a cada tasca específica. En aquesta tesi, la metodologia radiomics es va emprar per realitzar quatre projectes independents. En el primer projecte, vam estudiar la diferenciació entre glioblastomes multiformes (GBMs) i metàstasis cerebrals (MCs) en IRM convencional. Aquests tipus de tumors cerebrals poden confondre's al diagnosticar-se ja que presenten un perfil radiològic similar i les dades clíniques poden no ser concloents. Per tal d'evitar procediments exhaustius i invasius, vam estudiar el poder discriminatori de textures 2D extretes d'imatges de referència T1 filtrades i sense filtrar. Els resultats suggereixen que els paràmetres texturals proporcionen informació sobre l'heterogeneïtat dels GBMs i les MCs que pot servir per distingir amb precisió ambdues lesions quan s'utilitza una aproximació d'aprenentatge automàtic adequada. En el segon projecte, vam analitzar la classificació de MCs segons el seu origen primari en IRM de referència. En un percentatge de pacients, les MCs són diagnosticades com la primera manifestació d'un tumor primari desconegut. Per tal de detectar el tumor primari d'una forma no invasiva i més ràpida, vam examinar la capacitat de l'anàlisi de textura 2D i 3D per diferenciar les MCs derivades dels tumors primaris més propensos a metastatitzar (càncer de pulmó, càncer de mama i melanoma) en imatges T1. Els resultats van mostrar que s'aconsegueix una alta precisió quan s'utilitza un conjunt reduït de textures 3D per diferenciar les MCs de càncer de pulmó de les MCs de càncer de mama i melanoma. En el tercer projecte, vam avaluar les propietats de l'hipocamp en la IRM per identificar les diferents etapes de la malaltia d'Alzheimer (MA). Els criteris actuals per diagnosticar la MA requereixen la presència de dèficits cognitius severs. Amb la idea d'establir nous biomarcadors per detectar la MA en les seues primeres etapes, vam avaluar un conjunt de textures 2D i 3D extretes d'IRM de l'hipocamp de pacients amb MA avançada, deteriorament cognitiu lleu i normalitat cognitiva. Molts paràmetres de textura 3D van resultar ser estadísticament significatius per diferenciar entre pacients amb MA i individus de les altres dues poblacions. En combinar aquests paràmetres amb tècniques d'aprenentatge automàtic, es va obtenir una alta precisió. En el quart projecte, vam intentar caracteritzar els patrons d'heterogeneïtat de l'ictus cerebral isquèmic en la IRM estructural. En la IRM cerebral d'individus d'edat avançada, alguns processos patològics presenten característiques similars, com les lesions per ictus i les hiperintensitats de la substància blanca (HSBs). Atès que els ictus tenen efecte també en teixit adjacent, vam decidir estudiar la viabilitat de textures 3D extretes de les HSBs, la substància blanca no afectada i les estructures subcorticals per diferenciar individus afectats per ictus llacunars o corticals visibles en IRM convencional (imatges T1, T2 i FLAIR) d'individus sense ictus. Les textures no foren útils per diferenciar ictus corticals i llacunars, però es van obtenir resultats prometedors per disce
[EN] Over the last years, researchers have attempted to exploit the information provided by medical images through the evaluation of numerous imaging quantitative parameters in order to help clinicians with the diagnosis and assessment of many lesions and diseases. This practice has been recently named as radiomics. Texture analysis supply a wide range of features that allow quantifying the distinctive heterogeneity of different tissues, especially when obtained from magnetic resonance imaging (MRI). With this in mind, we decided to study the possibilities of texture features from MRI in order to characterize several disorders that affect the human brain. The potential of texture features was analyzed with various machine learning approaches, involving different classifiers and feature selection methods so as to find the optimal model to accomplish each specific task. In this thesis, the radiomics methodology was used to perform four independent projects. In the first project, we studied the differentiation between glioblastomas (GBMs) and brain metastases (BMs) in conventional MRI. Sometimes these types of brain tumors can be misdiagnosed since they may present a similar radiological profile and the clinical data may be inconclusive. With the aim of avoiding exhaustive and invasive procedures, we studied the discriminatory power of a large amount of 2D texture features extracted from baseline original and filtered T1-weighted images. The results suggest that 2D texture features provide some heterogeneity information of GBMs and BMs that can help in their accurate discernment when using the proper machine learning approach. In the second project, we analyzed the classification of BMs by their primary site of origin in baseline MRI. A percentage of patients are diagnosed with BM as the first manifestation of an unknown primary tumor. In order to detect the primary tumor in a faster non-invasive way, we examined the capability of 2D and 3D texture analysis to differentiate BMs derived from the most common primary tumors (lung cancer, breast cancer and melanoma) in T1-weighted images. The results showed that high accuracy was achieved when using a reduced set of 3D descriptors to differentiate lung cancer BMs from breast cancer and melanoma BMs. In the third project, we evaluated the hippocampus MRI profile of Alzheimer's disease (AD) patients to identify the different stages of the disease. The current criteria for diagnosing AD require the presence of relevant cognitive deficits. With the purpose of establishing new biomarkers to detect AD in its early stages, we evaluated a set of 2D and 3D texture features extracted from MRI scans of the hippocampus of patients with advanced AD, early mild cognitive impairment and cognitive normality. Many 3D texture parameters resulted to be statistically significant to differentiate between AD patients and subjects from the other two populations. When combining these 3D parameters with machine learning techniques, high accuracy was obtained. In the fourth project, we attempted to characterize the heterogeneity patterns of ischemic stroke in structural MRI. In brain MRI of older individuals, some pathological processes present similar imaging characteristics, like in the case of stroke lesions and white matter hyperintensities (WMH) of diverse natures. Given that stroke effects are present not only in the affected region, but also in unaffected tissue, we investigated the feasibility of 3D texture features from WMH, normal-appearing white matter and subcortical structures to differentiate individuals who had a lacunar or cortical stroke visible on conventional brain MRI (T1-weighted, T2-weighted and FLAIR images) from subjects who did not. Texture features were not useful to differentiate between post-acute cortical and lacunar strokes, but promising results were achieved for discerning between patients presenting an old stroke and normal-ageing patients who never had a stroke.
Ortiz Ramón, R. (2019). Radiomics for diagnosis and assessing brain diseases: an approach based on texture analysis on magnetic resonance imaging [Tesis doctoral no publicada]. Universitat Politècnica de València. https://doi.org/10.4995/Thesis/10251/119118
TESIS

Gliomas are considered to be the most common primary adult malignant brain tumor. With the dramatic increases in computational power and improvements in image analysis algorithms, computer-aided medical image analysis has been introduced into clinical applications. Precision tumor grading and genotyping play an indispensable role in clinical diagnosis, treatment and prognosis. Gliomas diagnostic procedures include histopathological imaging tests, molecular imaging scans and tumor grading. Pathologic review of tumor morphology in histologic sections is the traditional method for cancer classification and grading, yet human study has limitations that can result in low reproducibility and inter-observer agreement. Compared with histopathological images, Magnetic resonance (MR) imaging present the different structure and functional features, which might serve as noninvasive surrogates for tumor genotypes. Therefore, computer-aided image analysis has been adopted in clinical application, which might partially overcome these shortcomings due to its capacity to quantitatively and reproducibly measure multilevel features on multi-parametric medical information. Imaging features obtained from a single modal image do not fully represent the disease, so quantitative imaging features, including morphological, structural, cellular and molecular level features, derived from multi-modality medical images should be integrated into computer-aided medical image analysis. The image quality differentiation between multi-modality images is a challenge in the field of computer-aided medical image analysis. In this thesis, we aim to integrate the quantitative imaging data obtained from multiple modalities into mathematical models of tumor prediction response to achieve additional insights into practical predictive value. Our major contributions in this thesis are: 1. Firstly, to resolve the imaging quality difference and observer-dependent in histological image diagnosis, we proposed an automated machine-learning brain tumor-grading platform to investigate contributions of multi-parameters from multimodal data including imaging parameters or features from Whole Slide Images (WSI) and the proliferation marker KI-67. For each WSI, we extract both visual parameters such as morphology parameters and sub-visual parameters including first-order and second-order features. A quantitative interpretable machine learning approach (Local Interpretable Model-Agnostic Explanations) was followed to measure the contribution of features for single case. Most grading systems based on machine learning models are considered “black boxes,” whereas with this system the clinically trusted reasoning could be revealed. The quantitative analysis and explanation may assist clinicians to better understand the disease and accordingly to choose optimal treatments for improving clinical outcomes. 2. Based on the automated brain tumor-grading platform we propose, multimodal Magnetic Resonance Images (MRIs) have been introduced in our research. A new imaging–tissue correlation based approach called RA-PA-Thomics was proposed to predict the IDH genotype. Inspired by the concept of image fusion, we integrate multimodal MRIs and the scans of histopathological images for indirect, fast, and cost saving IDH genotyping. The proposed model has been verified by multiple evaluation criteria for the integrated data set and compared to the results in the prior art. The experimental data set includes public data sets and image information from two hospitals. Experimental results indicate that the model provided improves the accuracy of glioma grading and genotyping.

Le cancer du sein est une pathologie fréquente pour lequel les examens TEP/TDM au ¹⁸F-FDG et IRM mammaire sont fréquemment réalisés en routine. Il existe cependant une sous-utilisation des informations apportées par chacune de ces techniques d'imagerie. En pratique, l’interprétation de ces examens est principalement basée sur l’analyse visuelle et l'analyse « quantitative » se résume généralement au SUVmax seul en TEP/TDM et à l’étude du rehaussement du signal après injection de produit de contraste en IRM mammaire (DCE-MRI). L’arrivée de nouvelles machines hybrides TEP/ IRM, nous a amené à évaluer l'apport d’une quantification avancée des images issues de chacune de ces modalités séparément et en combinaison. Cela rejoint un domaine en expansion « la radiomique » qui consiste à extraire un grand nombre de caractéristiques quantitatives des images médicales pour décrypter l’hétérogénéité tumorale ou améliorer la prédiction du pronostic.L’objectif de notre travail était d’étudier l’apport des données radiomiques extraites de l’imagerie TEP au ¹⁸F-FDG et de l’IRM avec injection de produit de contraste réalisées avant traitement pour caractériser l’hétérogénéité tumorale dans le cancer du sein, en prenant en compte les différents sous-types moléculaires de cancer du sein, à savoir les tumeurs luminales (Lum A, Lum B HER2- et Lum B HER2+), triple-négatives et HER2+. Une importance particulière a été portée sur la valeur prédictive des informations radiomiques extraites de ces 2 techniques d’imagerie pour prédire le pronostic dans un groupe de patientes traitées par chimiothérapie néo-adjuvante. L’influence de variations physiologiques telles que l’âge sur le calcul des données radiomiques dans le tissu mammaire normal et cancéreux séparément a également été explorée, de même que la variabilité multicentrique des index radiomiques. L’extraction de ces données radiomiques a été effectuée grace au logiciel LiFex développé au sein du laboratoire IMIV sur une base de données-patientes recueillie en rétrospective.Nous avons rapporté pour la première fois l’influence de l’âge sur le calcul des indices « radiomiques » en TEP dans le tissu mammaire sain dans 2 institutions différentes mais aussi dans les tumeurs mammaires notamment celle triple-négatives. Des associations significatives entre le « phénotype tumoral radiomique » en imagerie TEP et IRM et des données pronostiques reconnues dans le cancer du sein ont été mises en évidence. En outre, nous avons démontré l’existence d’une grande variabilité pour le « profil radiomique » en TEP parmi les tumeurs présentant le même sous-type moléculaire. Cela suggére l’existence d’informations non-redondantes au sein du « phénotype tumoral métabolique » de chaque tumeur mammaire défini par les données radiomiques. L’exploration de cette variabilité s’est révélée intéressante pour améliorer la prédiction de la réponse histologique chez les patientes avec des tumeurs triple-négatives traitées par chimiothérapie néo-adjuvante. Par ailleurs, les mesures effectuées dans la région mammaire péri-tumorale chez les patientes traitées par chimiothérapie néo-adjuvante se sont montrées prédictives pour les patientes avec des tumeurs Lum B HER2-. En IRM nous avons montré l’importance de standardiser la méthode de mesure des caractéristiques radiomiques. Nous avons observé que les caractéristiques radiomiques issues des images DCE-MRI étaient moins associées aux caractéristiques moléculaires des tumeurs et avaient une valeur prédictive moindre. Nous avons également proposé une nouvelle méthode relativement standardisée pour le calcul des données radiomiques en IRM mammaire avec des résultats intéressants mais cette méthode doit encore être optimisée. Cependant, nos résultats suggèrent que les données extraites de la totalité du volume tumorale en IRM compléteraient efficacement les caractéristiques radiomiques TEP et le sous-type moléculaire pour prédire la réponse à la chimiothérapie néo-adjuvante
Breast cancer is a common disease for which ¹⁸F-FDG PET/CT and breast MRI are frequently performed in routine practice. However, the different information provided by each of these imaging techniques are currently under-exploited. Indeed, in routine the interpretation of these scans is mainly based on visual analysis whereas the « quantitative » analysis of PET/CT data is generally limited to the sole use of the SUVmax while in breast MRI, simple parameters to characterize tumor enhancement after injection of contrast medium are used. The advent of PET/MRI machines, calls for an evaluation of the contribution of a more advanced quantification of each of the modalities separately and in combination in the setting of breast cancer. This is along with the concept of « Radiomics » a field currently expanding and which consists in extracting many quantitative characteristics from medical images used in clinical practice to decipher tumor heterogeneity or improve prediction of prognosis. The aim of our work was to study the contribution of radiomic data extracted from ¹⁸F-FDG PET and MRI imaging with contrast injection to characterize tumor heterogeneity in breast cancer taking into account the different molecular subtypes of breast cancer, namely luminal (Lum A, Lum B HER2- and Lum B HER2 +), triple-negative and HER2 + tumors. In this context, we focused on the prediction of prognosis in patients treated with neo-adjuvant chemotherapy. The influence of physiological variations such as age on the calculation of radiomic data in normal breast and breast tumors separately was also explored, as well as the multi-center variability of radioman features. Radiomic features were extracted using the LiFex software developed within IMIV laboratory. The patient database used for the studies were all retrospective data. We reported for the first time the influence of age on the values of radiomic features in healthy breast tissue in patients recruited from 2 different institutions but also in breast tumors especially those with a triple-negative subtype. Similarly, significant associations between the radiomic tumor phenotype in PET and MRI imaging and well-established prognostic factors in breast cancer have been identified. In addition, we showed a large variability in the PET « radiomic profile » of breast tumors with similar breast cancer subtype suggesting complementary information within their metabolic phenotype defined by radiomic features. Moreover, taking into account this variability has been shown to be of particular interest in improving the prediction of pathological response in patients with triple-negative tumors treated with neoadjuvant chemotherapy. A peri-tumoral breast tissue region satellite to the breast tumor was also investigated and appeared to bear some prognostic information in patients with Lum B HER2- tumors treated with neoadjuvant chemotherapy. In MR, we demonstrated the need to harmonize the methods for radiomic feature calculation. Overall, we observed that radiomic features derived from MR were less informative about the molecular features of the tumors than radiomic features extracted from PET data and were of lower prognostic value. Yet, the combination of the enhanced tumor volume in MR with a PET radiomic feature and the tumor molecular subtype yielded enhanced the accuracy with which response to neoadjuvant therapy could be predicted compared to features from one modality only or molecular subtype only

ADVANCED IMAGING ANALYSIS FOR PREDICTING TUMOR RESPONSE AND IMPROVING CONTOUR DELINEATION UNCERTAINTY By Rebecca Nichole Mahon, MS A dissertation submitted in partial fulfillment of the requirements for the degree of Doctor of Philosophy at Virginia Commonwealth University. Virginia Commonwealth University, 2018 Major Director: Dr. Elisabeth Weiss, Professor, Department of Radiation Oncology Radiomics, an advanced form of imaging analysis, is a growing field of interest in medicine. Radiomics seeks to extract quantitative information from images through use of computer vision techniques to assist in improving treatment. Early prediction of treatment response is one way of improving overall patient care. This work seeks to explore the feasibility of building predictive models from radiomic texture features extracted from magnetic resonance (MR) and computed tomography (CT) images of lung cancer patients. First, repeatable primary tumor texture features from each imaging modality were identified to ensure a sufficient number of repeatable features existed for model development. Then a workflow was developed to build models to predict overall survival and local control using single modality and multi-modality radiomics features. The workflow was also applied to normal tissue contours as a control study. Multiple significant models were identified for the single modality MR- and CT-based models, while the multi-modality models were promising indicating exploration with a larger cohort is warranted. Another way advances in imaging analysis can be leveraged is in improving accuracy of contours. Unfortunately, the tumor can be close in appearance to normal tissue on medical images creating high uncertainty in the tumor boundary. As the entire defined target is treated, providing physicians with additional information when delineating the target volume can improve the accuracy of the contour and potentially reduce the amount of normal tissue incorporated into the contour. Convolution neural networks were developed and trained to identify the tumor interface with normal tissue and for one network to identify the tumor location. A mock tool was presented using the output of the network to provide the physician with the uncertainty in prediction of the interface type and the probability of the contour delineation uncertainty exceeding 5mm for the top three predictions.

Positron Emission Tomography (PET) is an imaging modality that has become increasingly beneficial in Radiotherapy by improving treatment planning (1). PET reveals tumor volumes that are not well visualized on computed tomography CT or MRI, recognizes metastatic disease, and assesses radiotherapy treatment (1). It also reveals areas of the tumor that are more radiosensitive allowing for dose painting - a non-homogenous dose treatment across the tumor (1). However, PET is not without limitations. The quantitative unit of PET images, the Standardized Uptake Value (SUV), is affected by many factors such as reconstruction algorithm, patient weight, and tracer uptake time (2). In fact, PET is so sensitive that a patient imaged twice in a single day on the same machine and same protocol will produce different SUV values. The objective of this research was to increase the capabilities of PET by exploring other quantitative PET/CT measures for Radiotherapy treatment applications. The technique of quantitative image feature analysis, nowadays known as radiomics, was applied to PET and CT images. Image features were then extracted from PET/CT images and how the features differed between conventional and respiratory-gated PET/CT images in lung cancer was analyzed. The influence of noise on image features was analyzed by applying uncorrelated, Gaussian noise to PET/CT images and measuring how significantly noise affected features. Quantitative PET/CT measures outside of image feature analysis were also investigated. The correlation of esophageal metabolic tumor volumes (tumor volume demonstrating high metabolic uptake) and endoscopically implanted fiducial markers was studied. It was found that certain image features differed greatly between conventional and respiratory-gated PET/CT. The differences were mainly due to the effect of respiratory motion including affine motion, rotational motion and tumor deformation. Also, certain feature groups were more affected by noise than others. For instance, contour-dependent shape features exhibited the least change with noise. Comparatively, GLSZM features exhibited the greatest change with added noise. Discordance was discovered between the inferior and superior tumor fiducial markers and metabolic tumor volume (MTV). This demonstrated a need for both fiducial markers and MTV to provide a comprehensive view of a tumor. These studies called attention to the differences in features caused by factors such as motion, acquisition parameters, and noise, etc. Investigators should be aware of these effects. PET/CT radiomic features are indeed highly affected by noise and motion. For accurate clinical use, these effects must be account by investigators and future clinical users. Further investigation is warranted towards the standardization of PET/CT radiomic feature acquisition and clinical application.

A CT-scan of lungs has become ubiquitous as a thoracic diagnostic tool. Thus, using CT-scan images in developing predictive models for tumor types and survival time of patients afflicted with Non-Small Cell Lung Cancer (NSCLC) would provide a novel approach to non-invasive tumor analysis. It can provide an alternative to histopathological techniques such as needle biopsy. Two major tumor analysis problems were addressed in course of this study, tumor type classification and survival time prediction. CT-scan images of 109 patients with NSCLC were used in this study. The first involved classifying tumor types into two major classes of non-small cell lung tumors, Adenocarcinoma and Squamous-cell Carcinoma, each constituting 30% of all lung tumors. In a first of its kind investigation, a large group of 2D and 3D image features, which were hypothesized to be useful, are evaluated for effectiveness in classifying the tumors. Classifiers including decision trees and support vector machines (SVM) were used along with feature selection techniques (wrappers and relief-F) to build models for tumor classification. Results show that over the large feature space for both 2D and 3D features it is possible to predict tumor classes with over 63% accuracy, showing new features may be of help. The accuracy achieved using 2D and 3D features is similar, with 3D easier to use. The tumor classification study was then extended by introducing the Bronchioalveolar Carcinoma (BAC) tumor type. Following up on the hypothesis that Bronchioalveolar Carcinoma is substantially different from other NSCLC tumor types, a two-class problem was created, where an attempt was made to differentiate BAC from the other two tumor types. To make a three-class problem a two-class problem, misclassification amongst Adenocarcinoma and Squamous-cell Carcinoma were ignored. Using the same prediction models as the previous study and just 3D image features, tumor classes were predicted with around 77% accuracy. The final study involved predicting two year survival time in patients suffering from NSCLC. Using a subset of the image features and a handful of clinical features, predictive models were developed to predict two year survival time in 95 NSCLC patients. A support vector machine classifier, naive Bayes classifier and decision tree classifier were used to develop the predictive models. Using the Area Under the Curve (AUC) as a performance metric, different models were developed and analyzed for their effectiveness in predicting survival time. A novel feature selection method to group features based on a correlation measure has been proposed in this work along with feature space reduction using principal component analysis. The parameters for the support vector machine were tuned using grid search. A model based on a combination of image and clinical features, achieved the best performance with an AUC of 0.69, using dimensionality reduction by means of principal component analysis along with grid search to tune the parameters of the SVM classifier. The study showed the effectiveness of a predominantly image feature space in predicting survival time. A comparison of the performance of the models from different classifiers also indicate SVMs consistently outperformed or matched the other two classifiers for this data.

Since the start of 2020 Sars-COVID19 has given rise to a world-wide pandemic. In an attempt to slow down the spreading of this disease various prevention and diagnostic methods have been developed. In this thesis the attention has been put on Machine Learning to predict prognosis based on data originating from radiological images. Radiomics has been used to extract information from images segmented using a software from the hospital which provided both the clinical data and images. The usefulness of different families of variables has then been evaluated through their performance in the methods used, i.e. Lasso regularized regression and Random Forest. The first chapter is introductory in nature, the second will contain a theoretical overview of the necessary concepts that will be needed throughout this whole work. The focus will be then shifted on methods and instruments used in the development of this thesis. The third chapter will report the results and finally some conclusions will be derived from the previously presented results. It will be concluded that the segmentation and feature extraction step is of pivotal importance in driving the performance of the predictions. In fact, in this thesis, it seems that the information from the images achieves the same predictive power that can be derived from the clinical data. This can be interpreted in three ways: first it can be taken as a symptom of the fact that even the more complex Sars-COVID19 cases can be segmented automatically, or semi-automatically by untrained personnel, leading to results competing with other methodologies. Secondly it can be taken to show that the performance of clinical variables can be reached by radiomic features alone in a semi-automatic pipeline, which could aid in reducing the workload imposed on medical professionals in case of pandemic. Finally it can be taken as proof that the method implemented has room to improve by more carefully investing in the segmentation phase

La survie globale des patients atteints de cancer du poumon non à petites cellules (CPNPC) métastatique a augmenté grâce à l’utilisation d’immunothérapies anti-PD1/PD-L1. Cependant, la durée de la réponse reste très variable d'un patient à l'autre, et seuls 20 à 30 % des patients sont encore en vie après deux ans. Par conséquent, de nouveaux biomarqueurs permettant de prédire la réponse au traitement et le pronostic des patients sont nécessaires pour guider la décision thérapeutique. Dans le cadre de mon doctorat, nous avons étudié des approches d'apprentissage automatique pour exploiter les données radiologiques, transcriptomiques et pathologiques, en les intégrant dans des modèles multimodaux susceptibles d'améliorer le pouvoir prédictif limité des données de routine clinique.Mon doctorat était au cœur d'un projet multidisciplinaire financé par la Fondation ARC, intitulé "SIGN'IT 2020-Signatures en Immunothérapie". Il réunissait plusieurs équipes de recherche de l'Institut Curie aux côtés d'une équipe de l'Institut du thorax, dirigée par le Professeur Nicolas Girard, en charge de la prise en charge des patients et de la collecte des données. Nous avons constitué une nouvelle cohorte multimodale de 317 patients atteints de CPNPC métastatique traités, en première ligne, par immunothérapie, seule ou associée à une chimiothérapie. Avant le début du traitement, nous avons recueilli des informations cliniques provenant des soins de routine, des examens TEP/TDM au 18F-FDG, des lames pathologiques numérisées provenant du diagnostic initial et des profils RNA-seq provenant de biopsies solides. Les résultats de l'immunothérapie ont été évalués en fonction de la survie globale (OS) et de la survie sans progression (PFS) de chaque patient.En collaboration avec Irène Buvat et Emmanuel Barillot, dont les équipes sont respectivement spécialisées dans l'analyse d'images médicales et de profils tumoraux RNA-seq, nous nous sommes d'abord concentrés sur la conception d'outils informatiques permettant d'extraire des informations pertinentes et interprétables à partir de ces deux modalités de données. Nous avons notamment développé un outil Python pour appliquer l'Analyse en Composantes Indépendantes (ICA) sur les données omiques et stabiliser les résultats à travers de multiples exécutions. Nous avons ensuite exploré le potentiel de l'ICA stabilisée pour extraire des caractéristiques transcriptomiques puissantes et biologiquement pertinentes pour la prédiction des résultats des patients. Pour les images médicales, et en particulier les examens TEP au 18F-FDG, nous avons étudié le potentiel des approches radiomiques pour caractériser la maladie métastatique au niveau du corps entier et concevoir de nouvelles caractéristiques prédictives. Nous avons conçu un outil d'explication Python, basé sur les valeurs de Shapley, pour mettre en évidence la contribution de chaque métastase individuelle à la prédiction des modèles radiomiques.Une part importante de mon doctorat a été consacrée à l'intégration des caractéristiques cliniques, radiomiques et transcriptomiques, ainsi que des caractéristiques pathomiques (avec l'aide de l'équipe de Thomas Walter). Nous avons procédé à une comparaison approfondie des capacités prédictives des différentes combinaisons multimodales en utilisant divers algorithmes d'apprentissage et méthodes d'intégration. Nous avons conçu des stratégies pour surmonter les nombreux défis associés à l'intégration multimodale, y compris la gestion des modalités manquantes pour de nombreux patients, la gestion d'une taille de cohorte modeste par rapport à la haute dimensionnalité des données, ou la garantie d'une comparaison équitable de toutes les combinaisons multimodales possibles. Nous nous sommes particulièrement attachés à mettre en évidence le potentiel des approches multimodales pour améliorer la stratification des risques des patients par rapport aux modèles utilisant uniquement des informations de routine clinique
Overall survival of patients with metastatic non-small cell lung cancer (NSCLC) has been increasing with the use of anti-PD-1 immune checkpoint inhibitors. However, the duration of response remains highly variable between patients, and only 20-30% of patients are alive at 2 years. Thus, new biomarkers for predicting response to treatment and patient outcomes are still needed to guide therapeutic decision. In my PhD, we investigated machine learning approaches to leverage radiological, transcriptomic, and pathological data, integrating them into powerful multimodal models that might improve the limited predictive power of routine clinical data.My doctoral research stood at the heart of a multidisciplinary project funded by Fondation ARC call «SIGN’IT 2020—Signatures in Immunotherapy». It brought together several research teams of Institut Curie alongside a team from Institut du thorax, led by Professor Nicolas Girard, in charge of patient management and data collection. We built a new multimodal cohort of 317 metastatic NSCLC patients treated with first-line immunotherapy alone or combined with chemotherapy. At baseline, we collected clinical information from routine care, 18F-FDG PET/CT scans, digitized pathological slides from the initial diagnosis, and bulk RNA-seq profiles from solid biopsies. Immunotherapy outcome was monitored with Overall Survival (OS) and Progression-Free Survival (PFS).Together with Irène Buvat and Emmanuel Barillot, whose teams hold significant expertise in the analysis of medical images and RNAseq tumor profiles, respectively, we initially focused on designing computational tools to extract relevant and interpretable information from these two data modalities. We notably developed a Python tool to apply Independent Component Analysis (ICA) on omics data and stabilize the results through multiple runs. We then explored the potential of stabilized ICA to extract powerful and biologically relevant transcriptomic features for the prediction of patient outcome. For medical images, and in particular 18F-FDG PET scans, we investigated the potential of radiomic approaches to characterize the metastatic disease at the whole-body level and design novel predictive features. We designed a Python explanation tool, based on Shapley values, to highlight the contribution of each individual metastasis to the prediction of radiomic models that use as input such whole-body features. A substantial portion of my PhD was devoted to the integration of clinical, radiomic, and transcriptomic features, as well as pathomic features extracted from digitized pathological slides (with the assistance of Thomas Walter’s team). We conducted a thorough comparison of the predictive capabilities of the different multimodal combinations using various state-of-the-art learning algorithms and integration methods. We devised strategies to overcome the many challenges associated to multimodal integration within our dataset, including handling missing modalities for numerous patients, dealing with a modest cohort size in comparison to the high dimensionality of the data, or ensuring a fair comparison of all the possible multimodal combinations. We especially focused on highlighting the potential of multimodal approaches to enhance patient risk stratification with respect to models using only clinical information collected during routine care

L'évolution de la texture ou de la forme d'une tumeur à l'imagerie médicale reflète les modifications internes dues à la progression (naturelle ou sous traitement) d'une lésion tumorale. Dans ces travaux nous avons souhaité étudier l'apport des caractéristiques delta-radiomiques pour prédire l'évolution de la maladie. Nous cherchons à fournir un pipeline complet de la reconstruction des lésions à la prédiction, en utilisant seulement les données obtenues en routine clinique.Tout d'abord, nous avons étudié un sous ensemble de marqueurs radiomiques calculés sur IRM, en cherchant à établir quelles conditions sont nécessaires pour assurer leur robustesse. Des jeux de données artificiels et cliniques nous permettent d'évaluer l'impact de la reconstruction 3D des zones d'intérêt et celui du traitement de l'image.Une première analyse d'un cas clinique met en évidence des descripteurs de texture statistiquement associés à la survie sans évènement de patients atteints d'un carcinome du canal anal dès le diagnostic.Dans un second temps, nous avons développé des modèles d'apprentissage statistique. Une seconde étude clinique révèle qu'une signature radiomique IRM en T2 à trois paramètres apprise par un modèle de forêts aléatoires donne des résultats prometteurs pour prédire la réponse histologique des sarcomes des tissus mous à la chimiothérapie néoadjuvante.Le pipeline d'apprentissage est ensuite testé sur un jeu de données de taille moyenne sans images, dans le but cette fois de prédire la rechute métastatique à court terme de patientes atteinte d'un cancer du sein. La classification des patientes est ensuite comparée à la prédiction du temps de rechute fournie par un modèle mécanistique de l'évolution des lésions.Enfin nous discutons de l'apport des techniques plus avancées de l'apprentissage statistique pour étendre l'automatisation de notre chaîne de traitement (segmentation automatique des tumeurs, analyse quantitative de l'oedème péri-tumoral)
Tumor shape and texture evolution may highlight internal modifications resulting from the progression of cancer. In this work, we want to study the contribution of delta-radiomics features to cancer-evolution prediction. Our goal is to provide a complete pipeline from the 3D reconstruction of the volume of interest to the prediction of its evolution, using routinely acquired data only.To this end, we first analyse a subset of MRI(-extracted) radiomics biomarquers in order to determine conditions that ensure their robustness. Then, we determine the prerequisites of features reliability and explore the impact of both reconstruction and image processing (rescaling, grey-level normalization). A first clinical study emphasizes some statistically-relevant MRI radiomics features associated with event-free survival in anal carcinoma.We then develop machine-learning models to improve our results.Radiomics and machine learning approaches were then combined in a study on high grade soft tissu sarcoma (STS). Combining Radiomics and machine-learning approaches in a study on high-grade soft tissue sarcoma, we find out that a T2-MRI delta-radiomic signature with only three features is enough to construct a classifier able to predict the STS histological response to neoadjuvant chemotherapy. Our ML pipeline is then trained and tested on a middle-size clinical dataset in order to predict early metastatic relapse of patients with breast cancer. This classification model is then compared to the relapsing time predicted by the mechanistic model.Finally we discuss the contribution of deep-learning techniques to extend our pipeline with tumor automatic segmentation or edema detection

La caractérisation non-invasive des gliomes fait partie de la médecine personalisée, aidant ainsi les cliniciens à prendre des décisions optimales pour améliorer la survie des patients tout en préservant leur qualité de vie. Dans ce contexte, l’imagerie moléculaire Tomographie par Emission de Positons (TEP) avec des radiotraceurs marqués aux acides aminés tels que la 18F-FDOPA, est actuellement recommandée par les groupes d’experts internationaux comme un complément à l’imagerie par résonance magnétique conventionnelle. Les progrès dans l’analyse d’images sont désormais axés sur la quantification de l’hétérogénéité tumorale par une analyse radiomique. Cependant, cette analyse a principalement été appliquée à des images statiques, en ignorant la dimension temporelle. Au contraire, l’analyse dynamique enregistre les variations temporelles du métabolisme tumoral, enrichissant ainsi l’évaluation statique. Malgré les résultats prometteurs des paramètres dynamiques à l’échelle de la région pour le diagnostic initial, leur efficacité dans la détection des récidives a été peu étudiée, avec certaines limitations identifiées pour cette indication. Cette thèse étudie donc l’analyse radiomique des images TEP dynamiques à l’échelle du voxel, en utilisant l’apprentissage automatique pour identifier des biomarqueurs d’intérêt dans les gliomes. La dimension temporelle de l’analyse radiomique a pu être étudiée à deux niveaux : en suivant la cinétique du métabolisme tumoral à partir d’une seule acquisition TEP dynamique, et en évaluant les variations métaboliques chez un même patient lors d’examens répétés dans le temps. Dans un premier temps, ce travail a évalué l’impact de l’application d’une fonction d’étalement du point sur l’analyse dynamique à l’échelle du voxel. Par la suite, une première analyse de la dimension temporelle a été étudiée avec une analyse radiomique effectuée à partir d’une seule image TEP dynamique. L’apport de cette analyse pour la détection des récidives de gliomes était faible. Par conséquent, la dimension temporelle de l’analyse radiomique a été réalisée en examinant les changements dans les caractéristiques radiomiques entre deux examens TEP consécutifs, chez des patients suivis pour gliome. Une étude de validation multicentrique a ensuite été effectuée pour évaluer l’apport de cette analyse radiomique en clinique. L’objectif était de déterminer si le modèle radiomique proposé permettait d’améliorer les performances diagnostiques des médecins dans la détection de l’agressivité des gliomes. Pour aller plus loin, l’adaptation des algorithmes d’apprentissage profond pour l’analyse de l’imagerie TEP à la 18F-FDOPA est encourageante. Cette approche pourrait offrir une flexibilité dans l’interprétabilité des modèles tout en évaluant les relations complexes entre les caractéristiques de l’imagerie TEP et le pronostic des patients
With the growing emphasis on personalized medicine, a non-invasive glioma characterization tool is essential, aiding clinicians in making optimal decisions to improve patient survival while preserving their quality of life. Medical imaging techniques such as Magnetic Resonance Imaging (MRI) and Positron Emission Tomography (PET) offer a promising solution in neuro-oncology for the non-invasive diagnosis and monitoring of gliomas. In this context, PET molecular imaging, particularly with amino acid radiotracers such as 18F-FDOPA, is currently recommended by international guidelines as an adjunct to conventional MRI. Advancements in image processing are now focused on quantifying tumor heterogeneity through the massive extraction of characteristics, known as radiomics analysis. However, this analysis has primarily been applied to static images acquired at a fixed time, ignoring the temporal dimension. In contrast, dynamic analysis offers a unique perspective by capturing the temporal variations of tumor metabolism, providing complementary information to static analysis. While region-based dynamic parameters have shown promising results for the initial diagnosis, they have limitations in detecting glioma recurrences. This thesis therefore explores the potential of machine learning-based radiomics analysis on dynamic PET acquisition at the voxel level to identify biomarkers of interest for glioma cancer indications. The temporal dimension of radiomics analysis can be addressed on two levels: by tracking the kinetics of tumor metabolism through single-time-point dynamic acquisition, and by monitoring changes in patient status over multiple examinations. Initially, this work investigated the impact of point spread function deconvolution, a common post-reconstruction technique at our institution, on voxel-based dynamic analysis. Subsequently, the first aspect of the temporal dimension was evaluated through radiomics analysis of single-time-point dynamic PET images at the voxel level. The prognostic value of this analysis for glioma recurrence detection was modest. Therefore, the temporal dimension of radiomics analysis was further explored by examining changes in radiomics features between two consecutive PET scans, aiming to monitor the post-treatment status of patients with glioma. A multicenter validation study was then conducted to assess the potential of integrating radiomics analysis into clinical practice. The objective was to investigate the impact of an explainable radiomics model on the diagnostic performance of physicians in determining the aggressiveness of suspected gliomas at the initial diagnosis. To go further, the feasibility of adapting deep learning algorithms to the analysis of 18F-FDOPA PET imaging is encouraging. This approach could provide greater flexibility in model explainability while capturing the complex relationships between PET imaging features and patient outcomes

Le diagnostic du gliome infiltrant du tronc cérébral (GITC) chez les enfants est l'un des plus éprouvants en oncologie pédiatrique. Malgré de nombreux essais cliniques explorant divers traitements, le pronostic reste sombre, la plupart des patients succombant entre 9 et 11 mois après le diagnostic. Les mutations génétiques clé associées au GITC incluent H3K27M, ACVR1 et TP53. Chaque mutation a des caractéristiques distinctes, poussant les médecins à suggérer des thérapies personnalisées, soulignant l'importance d'une détection précise des mutations pour guider le traitement. Situées dans la région cruciale du tronc cérébral, les tumeurs GITC présentent des risques significatifs liés à la biopsie en raison de potentiels dommages neurologiques. L'IRM est une méthode indispensable pour le diagnostic de ces tumeurs, évaluant leur extension et permettant de mesurer l'évolution de la maladie au cours de la thérapie. Une prédiction des mutations, combinée à l'identification des patients survivant plus de deux ans, pourrait améliorer la thérapie proposée à ces patients. Dans ce contexte, la radiomique transforme les images en vastes sources de données, extrayant des caractéristiques comme la forme et la texture pour aider à la prise de décision. L'objectif de cette thèse est de prédire les principales mutations génétiques et d'identifier les survivants à long terme, en mettant l'accent sur la normalisation des images et l'applicabilité des modèles radiomiques. Notre étude a utilisé une base de données rétrospective de l'Institut Gustave Roussy, comprenant les données IRM de 80 patients et leurs données cliniques respectives. Les données d'IRM ont mis en évidence des problèmes pour les études radiomiques, tels que l'inhomogénéité du champ de biais et l'effet "scanner". Pour répondre à ces défis, un pipeline de normalisation d'images IRM a été mis en place, et les caractéristiques radiomiques ont été harmonisées par la méthode ComBat. Pour faire face au problème de modalités manquantes dans l'ensemble de données, une stratégie multi-modèles a été employée, conduisant à 16 modèles distincts reposant sur diverses combinaisons de caractéristiques radiomiques et cliniques. Cette approche a ensuite été rationalisée en une méthode multimodale, réduisant le nombre de modèles à cinq, après une phase de sélection de caractéristiques indépendantes. Les résultats de l'approche multimodale se sont avérés être prometteurs. Cette stratégie multimodale a été essentielle pour identifier les patients survivant plus de deux ans et a été complétée par l'approche ICARE pour une analyse de survie détaillée
The diagnosis of diffuse intrinsic pontine glioma (DIPG) in children stands as one of the most harrowing within pediatric oncology. Despite numerous clinical trials exploring various treatments, the prognosis remains bleak, with most patients succumbing between 9 to 11 months post-diagnosis. Key gene mutations linked to DIPG include H3K27M, ACVR1, and TP53. Each mutation has distinct characteristics, leading physicians to suggest tailored therapies, underscoring the importance of accurate mutation detection in guiding treatment. Located in the crucial region of the brainstem, the pons, DIPG tumors pose significant biopsy risks due to potential neurological damage. Hence, MRI could become a primordial diagnostic tool for these tumors, assessing their spread and gauging therapy responses. Its use to predict accurate gene mutation, and identify long-term survivors, could enhance patient care significantly. Within this framework, radiomics transforms images into vast data sources, extracting features like shape and texture to aid decision-making. The objective of this thesis is to refine mutation prediction and pinpoint long-term survivors, emphasizing image normalization and the applicability of radiomic models. Our study utilized a retrospective database from Gustave Roussy Institute, encompassing 80 patients MRI data and their respective clinical data. These MRI images highlighted issues in radiomic studies, such as bias field inhomogeneity and the "scanner effect". To address these challenges, a dedicated MR image normalization pipeline was implemented, and radiomic features underwent ComBat harmonization. Given the dataset's missing modalities, a multi-model strategy was employed, leading to 16 distinct models based on various radiomic and clinical feature combinations. This approach was then streamlined into a multi-modal method, reducing the number of models to five. The results from the ensemble of these models proved to be the most promising. This multi-modal strategy incorporated a feature selection phase, pinpointing the most pertinent features. Additionally, this method was applied to identify long-term survivors and was complemented by the ICARE framework for a nuanced survival analysis output

Introduction There are several radiological techniques, and, in this study, we focused the Single Photon Emission Computed Tomography (SPECT) imaging. It is possible to reconstruct the unknown tracer distribution inside the body by applying tomographic reconstruction algorithms such as Filtered Back Projection (FBP) and Ordered Subset Expectation Maximisation (OSEM) to the acquired data. Nowadays, thanks to technological innovations, a new branch of research has rapidly evolved: the radiomics. In practice, radiomics tries to assess whether the “textural features” of images in regions related to specific diseases could provide added value in a diagnostic process, in the evaluation of prognosis or could guide therapeutic choices. A general concern that has to be accounted for when performing a clinical study is whether changes in acquisition and reconstruction parameters could affect the value of the features. To the best of our knowledge, in literature, there is no unique method for identifying robust features; here, a generalised method to study the effects of the variation of reconstruction parameters on radiomic features is proposed and applied to asses stability and reliability in SPECT imaging. Materials and methods Only simulation studies could asses the link between features extracted from reconstructed images and their original values. From a preliminary statistical analysis, it emerged that at least 66 phantoms (representing different original textures) were needed to achieve a statistical power higher than 90%. These synthetic phantoms derived from abdominal CT scans and “Visible Human Project” image sets. Then, using a proper model, we simulated SPECT acquisitions of each phantom and reconstructed the corresponding images changing parameters in FBP and OSEM tomographic algorithms. Features extraction was conducted with PyRadiomics, an open-source software. Six feature classes were considered, based on Intensity, Grey-Level Co-occurrence Matrix (GLCM), Grey Level Dependence Matrix (GLDM), Grey Level Run Length Matrix (GLRLM), Grey Level Size Zone Matrix (GLSZM) and Neighbourhood Grey-Tone Difference Matrix (NGTDM). Ultimately, 93 different radiomics features for each phantom were calculated. In this way, data-set has a series of repeated measurements and the method of Generalised Estimating Equations (GEE) is suitable for analysing databases with a similar structure. In this study, two different GEE models were developed: one to analyse if the radiomic features calculated in the reconstructed images (Vr) reproduce the same feature in the original VOI (Vo); another to study if they are stable or not with reconstruction parameters variations. Results 32 different reconstructions for each available phantom were obtained, for a total of 2112 images stacks. The results of the two GEE models, features could be classified according to four possible groups: a) feature with a correlation between Vo and Vr, without reconstruction parameters variation effect; b) features without a correlation between Vo and Vr and without a significant impact of the reconstruction parameters variation; c) features with a statistically significant correlation between Vo and Vr and with the effect of the reconstruction parameters variation on the Vr value; d) reconstruction parameters variation affects Vr. Moreover, there is not a correlation between the values obtained from the reconstructed images and Vo. Discussion In literature, as far as we know, there are no trustworthy works of the reproducibility or repeatability applied to SPECT imaging. Here, with software simulations, we tried to answer the following two questions: 1) Are the features extracted from the reconstructed images (Vr) correlated to those of the original images (Vo)? 2) Are the features extracted from the reconstructed images robust when the reconstruction parameters vary? To answer these questions, two GEE models were developed. Most features showed a correlation between Vo and Vr, but with a relevant impact of reconstruction parameters variation. For clinical studies, in our opinion, features like a) would be the optimal choice. However, also features like c) could be used, but researchers have to handle with care these features for which the reconstruction parameters variations affect Vr. Using the remaining features is not recommended as the lack of correlation between Vo and Vr makes random any link with clinical end-points, so it could be difficult to reproduce any result on cohorts of patients other than the one used to develop the radiomic model. Conclusions From this study, it emerges how reconstruction parameters could affect radiomic features in SPECT imaging. In our opinion, researchers should take into account this dependency in both retrospective and prospective radiomic studies. Ultimately, the method described in this work, although complicated, represents a logical approach to carry out propaedeutic evaluations about the selection of imaging parameters or radiomic features to be used for clinical studies.

Radiomics treats images as quantitative data and promises to improve cancer prediction in radiology and therapy response assessment in radiation oncology. However, there are a number of fundamental problems that need to be solved in order to potentially apply radiomic features in clinic. The first basic step in computed tomography (CT) radiomic analysis is the acquisition of images using selectable image acquisition and reconstruction parameters. Radiomic features have shown large variability due to variation of these parameters. Therefore, it is important to develop methods to address these variability issues in radiomic features due to each CT parameter. To this end, texture phantoms provide a stable geometry and Hounsfield Units (HU) to characterize the radiomic features with respect to image acquisition and reconstruction parameters. In this project, normalization methods were developed to address the variability issues in CT Radiomics using texture phantoms. In the first part of this project, variability in radiomic features due to voxel size variation was addressed. A voxel size resampling method is presented as a preprocessing step for imaging data acquired with variable voxel sizes. After resampling, variability due to variable voxel size in 42 radiomic features was reduced significantly. Voxel size normalization is presented to address the intrinsic dependence of some key radiomic features. After normalization, 10 features became robust as a function of voxel size. Some of these features were identified as predictive biomarkers in diagnostic imaging or useful in response assessment in radiation therapy. However, these key features were found to be intrinsically dependent on voxel size (which also implies dependence on lesion volume). The normalization factors are also developed to address the intrinsic dependence of texture features on the number of gray levels. After normalization, the variability due to gray levels in 17 texture features was reduced significantly. In the second part of the project, voxel size and gray level (GL) normalizations developed based on phantom studies, were tested on the actual lung cancer tumors. Eighteen patients with non-small cell lung cancer of varying tumor volumes were studied and compared with phantom scans acquired on 8 different CT scanners. Eight out of 10 features showed high (Rs > 0.9) and low (Rs < 0.5) Spearman rank correlations with voxel size before and after normalizations, respectively. Likewise, texture features were unstable (ICC < 0.6) and highly stable (ICC > 0.9) before and after gray level normalizations, respectively. This work showed that voxel size and GL normalizations derived from texture phantom also apply to lung cancer tumors. This work highlights the importance and utility of investigating the robustness of CT radiomic features using CT texture phantoms. Another contribution of this work is to develop correction factors to address the variability issues in radiomic features due to reconstruction kernels. Reconstruction kernels and tube current contribute to noise texture in CT. Most of texture features were sensitive to correlated noise texture due to reconstruction kernels. In this work, noise power spectra (NPS) was measured on 5 CT scanners using standard ACR phantom to quantify the correlated noise texture. The variability in texture features due to different kernels was reduced by applying the NPS peak frequency and the region of interest (ROI) maximum intensity as correction factors. Most texture features were radiation dose independent but were strongly kernel dependent, which is demonstrated by a significant shift in NPS peak frequency among kernels. Percent improvements in robustness of 19 features were in the range of 30% to 78% after corrections. In conclusion, most texture features are sensitive to imaging parameters such as reconstruction kernels, reconstruction Field of View (FOV), and slice thickness. All reconstruction parameters contribute to inherent noise in CT images. The problem can be partly solved by quantifying noise texture in CT radiomics using a texture phantom and an ACR phantom. Texture phantoms should be a pre-requisite to patient studies as they provide stable geometry and HU distribution to characterize the radiomic features and provide ground truths for multi-institutional validation studies.

One major objective in radiation oncology is the personalisation of cancer treatment. The implementation of this concept requires the identification of biomarkers, which precisely predict therapy outcome. Besides molecular characterisation of tumours, a new approach known as radiomics aims to characterise tumours using imaging data. In the context of the presented thesis, radiomics was established at OncoRay to improve the performance of imaging-based risk models. Two software-based frameworks were developed for image feature computation and risk model construction. A novel data-driven approach for the correction of intensity non-uniformity in magnetic resonance imaging data was evolved to improve image quality prior to feature computation. Further, different feature selection methods and machine learning algorithms for time-to-event survival data were evaluated to identify suitable algorithms for radiomics risk modelling. An improved model performance could be demonstrated using computed tomography data, which were acquired during the course of treatment. Subsequently tumour sub-volumes were analysed and it was shown that the tumour rim contains the most relevant prognostic information compared to the corresponding core. The incorporation of such spatial diversity information is a promising way to improve the performance of risk models.:1. Introduction 2. Theoretical background 2.1. Basic physical principles of image modalities 2.1.1. Computed tomography 2.1.2. Magnetic resonance imaging 2.2. Basic principles of survival analyses 2.2.1. Semi-parametric survival models 2.2.2. Full-parametric survival models 2.3. Radiomics risk modelling 2.3.1. Feature computation framework 2.3.2. Risk modelling framework 2.4. Performance assessments 2.5. Feature selection methods and machine learning algorithms 2.5.1. Feature selection methods 2.5.2. Machine learning algorithms 3. A physical correction model for automatic correction of intensity non-uniformity in magnetic resonance imaging 3.1. Intensity non-uniformity correction methods 3.2. Physical correction model 3.2.1. Correction strategy and model definition 3.2.2. Model parameter constraints 3.3. Experiments 3.3.1. Phantom and simulated brain data set 3.3.2. Clinical brain data set 3.3.3. Abdominal data set 3.4. Summary and discussion 4. Comparison of feature selection methods and machine learning algorithms for radiomics time-to-event survival models 4.1. Motivation 4.2. Patient cohort and experimental design 4.2.1. Characteristics of patient cohort 4.2.2. Experimental design 4.3. Results of feature selection methods and machine learning algorithms evaluation 4.4. Summary and discussion 5. Characterisation of tumour phenotype using computed tomography imaging during treatment 5.1. Motivation 5.2. Patient cohort and experimental design 5.2.1. Characteristics of patient cohort 5.2.2. Experimental design 5.3. Results of computed tomography imaging during treatment 5.4. Summary and discussion 6. Tumour phenotype characterisation using tumour sub-volumes 6.1. Motivation 6.2. Patient cohort and experimental design 6.2.1. Characteristics of patient cohorts 6.2.2. Experimental design 6.3. Results of tumour sub-volumes evaluation 6.4. Summary and discussion 7. Summary and further perspectives 8. Zusammenfassung

The purpose of this Ph.D. thesis is to illustrate the research works carried out during the conceptualization, design, implementation, and evaluation of novel Clinical Decision Support Systems (CDSSs) based on Radiomics, Pathomics and Deep Learning (DL) techniques. CDSSs can be effective systems for implementing Precision Medicine into clinical practice since they permit the objective and repeatable evaluation of patients. Precision Medicine can enable the improvement of the healthcare system by employing a personal healthcare process for the health status of an individual patient, which evolves in a unique way. The methodologies concerning CDSSs were developed with different underlying goals: improvement of the clinical results, availability and usability of the method, and feasibility of the integration into the routine clinical practice. The applications considered span from Radiology to Digital Pathology. Tasks under consideration in Medical Imaging applications, from a computer vision perspective, concerned object detection, instance segmentation, semantic segmentation, color normalization, and characterization and classification of regions of interest. Data under consideration were either provided by local hospitals or obtained from public repositories. Validation of the developed systems has been done in accordance with the physicians. Moreover, the explainability of the realized systems has been investigated, by analyzing features' structure or by means of perceptive saliency maps. In the aforementioned scenario, the main purpose of this thesis is to develop new systems based on Deep Learning, Radiomics and Pathomics for the processing and analysis of medical images. Computational Imaging is a promising methodology to incorporate in the framework of Precision Medicine. Indeed, it creates the possibility to characterize the lesions in large datasets of images belonging to Radiology and Digital Pathology domains in an effective way, offering a personalized evaluation of the patient. Merits and shortcomings regarding DL in the field of Medical Imaging have been investigated for applications in Radiology and Digital Pathology. Technical contributions include devising novel algorithms, improving existing workflows, and assembling complex CDSSs by combining in an original and effective way different techniques proposed in the literature. In the Radiology domain, the following tasks have been tackled for what concerns applications related to Image-guided Surgery (IGS): liver segmentation, including also the classification into anatomical segments; vertebrae segmentation and identification; prostate segmentation and registration for image fusion. Radiomics has been exploited for characterizing lung lesions in COVID-19 patients, in order to discover a prognostic signature for those with a higher risk of developing pulmonary thromboembolism. With regard to Digital Pathology, applications included colorectal cancer (CRC) tissue classification; hematoxylin and eosin (H&E) stain color normalization; nuclei segmentation and detection; glomeruli lesions classifications according to Oxford score for IgA nephropathy patients. These automatic pipelines for histological data analysis can enable Pathomics, allowing the objective quantification and evaluation of tissue patterns. The developed solutions in all these scenarios were put in comparison with state-of-the-art approaches proposed in the literature, and were validated with physicians when possible. In many cases, data have also been collected from local institutions. This thesis work is organized into five chapters. Chapter 1 introduces the objective and the technical contribution of the thesis. Chapter 2 describes the state-of-the-art in all the considered clinical scenarios, with a particular focus on Radiology and Digital Pathology, encompassing emerging trends such as Radiomics and Pathomics. Chapter 3 describes the contributions proposed in the Radiology field. In particular, IGS applications concern liver segmentation and classification into segments, vertebrae segmentation and identification, and prostate segmentation and registration. Also, a Radiomics-based analysis of lung lesions of patients diagnosed with COVID-19 is presented. Chapter 4 presents the contributions proposed in the field of Digital Pathology, concerning tissue segmentation, normalization and classification, and detection of objects of interest, such as nuclei of cells. Lastly, final remarks and considerations for future works are drawn in Chapter 5.

abstract: The rate of progress in improving survival of patients with solid tumors is slow due to late stage diagnosis and poor tumor characterization processes that fail to effectively reflect the nature of tumor before treatment or the subsequent change in its dynamics because of treatment. Further advancement of targeted therapies relies on advancements in biomarker research. In the context of solid tumors, bio-specimen samples such as biopsies serve as the main source of biomarkers used in the treatment and monitoring of cancer, even though biopsy samples are susceptible to sampling error and more importantly, are local and offer a narrow temporal scope. Because of its established role in cancer care and its non-invasive nature imaging offers the potential to complement the findings of cancer biology. Over the past decade, a compelling body of literature has emerged suggesting a more pivotal role for imaging in the diagnosis, prognosis, and monitoring of diseases. These advances have facilitated the rise of an emerging practice known as Radiomics: the extraction and analysis of large numbers of quantitative features from medical images to improve disease characterization and prediction of outcome. It has been suggested that radiomics can contribute to biomarker discovery by detecting imaging traits that are complementary or interchangeable with other markers. This thesis seeks further advancement of imaging biomarker discovery. This research unfolds over two aims: I) developing a comprehensive methodological pipeline for converting diagnostic imaging data into mineable sources of information, and II) investigating the utility of imaging data in clinical diagnostic applications. Four validation studies were conducted using the radiomics pipeline developed in aim I. These studies had the following goals: (1 distinguishing between benign and malignant head and neck lesions (2) differentiating benign and malignant breast cancers, (3) predicting the status of Human Papillomavirus in head and neck cancers, and (4) predicting neuropsychological performances as they relate to Alzheimer’s disease progression. The long-term objective of this thesis is to improve patient outcome and survival by facilitating incorporation of routine care imaging data into decision making processes.
Dissertation/Thesis
Doctoral Dissertation Biomedical Informatics 2017