Machine Learning for predicting Immunotherapy response from routine Medical Imaging

Abstract

[eng] HYPOTHESIS: Precision oncology has revolutionized the landscape of cancer treatment. Emerging therapeu- tic approaches, such as immunotherapy or anti-angiogenic agents, have exhibited remarkable outcomes in solid tumors. Despite the comprehensive understanding of the underlying bio- logical mechanism driving these techniques, the translation of targeted therapies into clinical practice have resulted in benefits for only a subset of patients. Efforts have been dedicated to develop biomarkers for predicting treatment outcomes and improving patient selection. While significant progress has been made in tailoring therapeutic strategies to specific molecular and biological subtypes, no robust biomarker obtained to date can accurately anticipate tumor response. The general hypothesis within this thesis is founded on the premise that extracting robust quantitative data from standard medical images, we can effectively capture tumor pheno- types. These phenotypes could serve as indicators to identify patients who are more likely to respond to immunotherapy and other targeted therapies, thereby advancing the field of precision oncology. Specifically, we hypothesize that: • Robust radiomic features, accounting for multi-centric acquisition and standardized preprocessing, will enhance the performance of predictive models. • It is possible to capture patterns of response to immunotherapy from medical imaging representations of tumor phenotypes. • Application of radiomic predictive models can generalize to other imaging modalities and specific treatment responses. • Response to immunotherapy can be predicted from different sources of medical imaging, such as radiological and pathological imaging.

OBJECTIVES: The general objective of this thesis is to evaluate the potential of quantitative medical image and machine learning modeling in deriving accurate tumor phenotype representations for patient stratification in cancer treatment. More specifically, the objectives of this thesis are: • Objective 1: To identify potential sources of variability in CT imaging acquisition that could affect Radiomics reproducibility and explore methods for variability correc- tion. • Objective 2: To develop and validate radiomic signatures derived from CT scans for predicting patient response in a population with different tumor types treated with immunotherapy. • Objective 3: To explore the use of radiomics in anti-angiogenics targeted therapies and PET imaging. • Objective 4: To investigate the performance of DL-methods applied to histological images to predict respon