Thèse Développement de Méthodes d'Intelligence Artificielle Multimodale pour l'Analyse de l'Irm dans le Cancer de l'Ovaire Avancé H/F - 34

Établissement : Université de Montpellier
École doctorale : I2S - Information, Structures, Systèmes
Laboratoire de recherche : IRCM - Institut de Recherche en Cancérologie de Montpellier
Direction de la thèse : Stéphanie NOUGARET ORCID 0000000263988648
Début de la thèse : 2026-10-01
Date limite de candidature : 2026-04-30T23:59:59

Le cancer ovarien séreux de haut grade est caractérisé par une forte hétérogénéité tumorale et un taux élevé de rechute après chimiothérapie à base de sels de platine. La résistance au platine, définie par une récidive précoce après traitement, constitue le principal déterminant du pronostic. Malgré l'identification de facteurs moléculaires tels que les mutations BRCA ou le statut HRD, la capacité à prédire précocement la résistance reste limitée.

Ce projet de thèse vise à développer et valider des modèles prédictifs multimodaux de résistance au platine à partir d'une large cohorte multicentrique sino-française d'environ 600 patientes atteintes de cancer ovarien avancé. Cette cohorte comprend des IRM pelviennes standardisées, des données cliniques complètes (FIGO, résécabilité, KELIM), des données mutationnelles, des informations histopathologiques et un suivi de survie à long terme.

La première étape consistera à mettre en place une segmentation automatique par apprentissage profond des masses ovariennes et des lésions de carcinose péritonéale sur IRM. Cette segmentation permettra une quantification volumétrique précise de la charge tumorale ainsi qu'une analyse spatiale détaillée.

Dans un second temps, des caractéristiques radiomiques conformes aux standards IBSI seront extraites des volumes segmentés, incluant des descripteurs de texture, de forme et d'intensité. Une attention particulière sera portée à l'évaluation de l'hétérogénéité intratumorale, notamment via des approches de clustering spatial et de mesure d'entropie, afin de caractériser la structure interne des tumeurs et des implants péritonéaux.

L'objectif central est d'identifier des signatures radiomiques associées à la résistance au platine et à la survie sans progression. Ces signatures seront intégrées dans des modèles prédictifs multimodaux combinant imagerie, données cliniques (dont le score KELIM), statut mutationnel et paramètres histologiques.

Des approches statistiques pénalisées et des méthodes d'apprentissage automatique seront utilisées afin de limiter le surapprentissage et d'assurer la robustesse des modèles. La cohorte multicentrique permettra une validation interne croisée et une validation externe entre centres.

Ce projet ambitionne de dépasser les approches basées uniquement sur des facteurs cliniques ou moléculaires en intégrant la dimension spatiale de l'hétérogénéité tumorale visible en IRM. Il contribuera à une meilleure stratification des patientes dès le diagnostic et pourrait guider des stratégies thérapeutiques adaptées aux profils de risque.

High-grade serous ovarian cancer (HGSOC) remains one of the most lethal gynecological malignancies, largely due to late diagnosis and high relapse rates after platinum-based chemotherapy. Although the majority of patients initially respond to treatment, a substantial proportion develop platinum resistance, defined by early recurrence following platinum exposure. Platinum resistance represents the principal determinant of long-term survival and therapeutic limitation.

Current prognostic stratification relies on clinical parameters (FIGO stage, residual disease), biological markers such as BRCA mutations and homologous recombination deficiency (HRD), and dynamic response indicators such as the KELIM score. However, these variables fail to capture the spatial organization and heterogeneity of tumor tissue, which are increasingly recognized as central drivers of therapeutic adaptation and resistance.

Magnetic resonance imaging (MRI) provides a non-invasive representation of tumor morphology and microstructural properties. Radiomics enables high-dimensional quantitative feature extraction from medical images, potentially revealing latent phenotypic patterns invisible to conventional visual interpretation. Several studies have suggested associations between radiomic features and treatment response in ovarian cancer, but most have been limited by small sample sizes, single-center design, and lack of external validation. Moreover, few studies have rigorously addressed multicentric variability, domain shift, and model generalizability.

This thesis project leverages a large multicentric Sino-French cohort of approximately 600 patients with advanced ovarian cancer, including standardized MRI examinations, clinical data, mutation profiles, histopathological information, and survival outcomes. The scale and diversity of this dataset provide a unique opportunity to develop robust, generalizable predictive models of platinum resistance grounded in quantitative imaging and computational modeling.

The overarching objective of this thesis is to develop and validate reproducible, interpretable, and generalizable multimodal predictive models of platinum resistance in advanced ovarian cancer using large-scale multicentric MRI data.

Specific objectives are:

To design and validate automated deep-learning-based segmentation algorithms for primary ovarian tumors and peritoneal carcinomatosis on MRI across multiple centers.

To quantitatively characterize intratumoral and inter-lesional heterogeneity using advanced radiomic and spatial modeling approaches.

To identify imaging-derived signatures predictive of platinum resistance and progression-free survival.

To develop multimodal predictive models integrating radiomic features with clinical variables (including KELIM), mutational status (BRCA/HRD), and pathological parameters.

To evaluate model robustness, stability, and generalizability across centers through rigorous validation strategies.

To implement interpretable machine learning approaches enabling identification of biologically and clinically meaningful predictive features.

The methodological framework combines image processing, radiomics, statistical modeling, and machine learning under a robust validation strategy.

First, automated tumor segmentation will be implemented using convolutional neural networks based on U-Net or transformer-enhanced architectures. The models will be trained on manually annotated datasets and validated using Dice similarity coefficients, Hausdorff distance, and cross-center reproducibility metrics. Domain adaptation strategies, including intensity normalization and adversarial training approaches, will be explored to mitigate inter-site variability.

Following segmentation, radiomic features will be extracted in compliance with IBSI standards. Features will include first-order statistics, texture matrices (GLCM, GLRLM, GLSZM), neighborhood gray-tone difference metrics, wavelet decompositions, and morphological descriptors. Feature robustness will be evaluated through test-retest analysis where available and stability analysis under segmentation perturbations.

To characterize heterogeneity, spatial clustering methods such as k-means, Gaussian mixture models, and graph-based clustering will be applied to voxel-wise feature maps. Entropy-based metrics, habitat proportions, and spatial interface complexity will be computed to quantify tumor organization. These heterogeneity descriptors will be analyzed both globally and separately for primary tumors and peritoneal implants.

Dimensionality reduction will be performed using correlation filtering, principal component analysis, and penalized regression (LASSO/elastic net). The primary endpoint will be platinum resistance classification. Secondary endpoints will include progression-free survival and overall survival.

Predictive modeling will compare several machine learning approaches, including penalized logistic regression, random forests, gradient boosting machines, and support vector machines. Nested cross-validation will be implemented to prevent information leakage. External validation will be performed through inter-center training-testing splits, ensuring model generalizability across geographic populations.

Multimodal integration will incorporate clinical variables (including KELIM score), mutational status (BRCA/HRD), and histopathological features. Feature importance will be assessed using SHAP values to ensure interpretability. Model performance will be evaluated using AUC, sensitivity, specificity, calibration curves, Brier scores, and decision-curve analysis.

All computational pipelines will be developed in reproducible frameworks with version-controlled code and standardized preprocessing to facilitate transparency and scalability.