Responsible for 17,000 deaths per year in France, colorectal cancer ranks third among the most common cancers in both men and women. Diagnosed at a localized stage, it is easily treated with surgery. However, if it spreads and becomes metastatic, chemotherapy remains the standard treatment. Unfortunately, these cancers have not benefited from two major advances in oncology: targeted therapies and immunotherapies. To address this therapeutic impasse, the project is deploying a predictive diagnostic system for treatments, including immunotherapies, to detect effective treatments, guide them as early as possible, and thus maximize the chances of recovery. On this axis biomathematics develop novel image-based methodologies for the processing of videos from organoids towards personalize treatment.
Announcement: https://www.gustaveroussy.fr/fr/rhu-organomic-des-organoides-prediction-traitements-cancers-digestifs
Project Page: https://rhu-organomic.com/
Breast cancer stands as the most prevalent cancer worldwide, with WHO data revealing 2.3 million new cases, 685,000 deaths in 2020, and 7.8 million women living with a recent diagnosis in the past 5 years. Appropriate patient selection strategies can, therefore, greatly impact the treatment outcomes and patient life expectancy. Histopathology tissue slides are routinely used for tumor phenotyping and treatment selection, where experienced pathologists analyze these slides with a microscope to quantify different biomarkers highlighted under a variety of tissue staining protocols. Lately, these tissue slides have been digitized into gigapixel size whole slide images (WSIs), enabling automatic processing and promoting research to discover biological processes. Nevertheless, processing such large images presents significant challenges for traditional techniques. Deep learning (DL) has emerged as a powerful technique, driving progress in many visual recognition tasks across a broad span of disciplines. Despite the remarkable performance achieved by DL in understanding medical images, including histology, several important drawbacks limit its deployment in real-life scenarios, which could affect the lives of millions of people. First, modern DL models do not generalize well to unseen scenarios, requiring large labeled training datasets for each novel task. Second, recent evidence exposes that high-capacity networks suffer from poor calibration, producing overconfident estimates even in situations of high uncertainty, leading to unreliable models. It is therefore desired to empower learning strategies that i) accommodate DL models to new tasks with a few labeled samples, ii) in a computationally efficient manner, and iii) abstain model decisions from unreliable predictions.
Artificial Intelligence and its current success in a variety of fields influences also considerably the field of health care. A great number of algorithms currently provide powerful solutions, assisting medical doctors in their everyday practice in particular for diagnostic purposes. The ambition of this grant proposal is to harness the latest advances and to propose novel, explainable, and computationally efficient artificial intelligence algorithms and companion biomarkers that go beyond diagnosis and address patient stratification, treatment selection, and prognosis for lung cancer that is among the most lethal and associated with low survival. The proposed project is highly interdisciplinary, as it lies at the interface between clinical aspects, such as tools for patient stratification or treatment compatibility, and technical aspects, such as machine and deep learning for explainable representation of medical data and ensemble learning for multi-omics analysis. Highly accurate and robust algorithms are expected to extract valuable information from different biomedical data that can influence both medical and technical communities towards precision medicine.
URL: https://anr.fr/Project-ANR-21-CE45-0007
Cancer is the leading cause of death worldwide. Lately, precision medicine efforts are shifting the cancer care standards providing novel personalized treatment plans with promising outcomes. Patient selection for such treatment regimes is principally based on the assessment of chemically stained histopathological tissue slides that can provide a distinct readout of different molecular profiles such as KRAS, HER2, Ki67 and others. Additionally, evidence highlight the importance of the spatial distribution of such molecular phenotypes for treatment decisions. Within the scope of this project, we aim to design and evaluate novel artificial intelligence models that computationally quantify spatially resolved treatment associated information from digital whole slide images. Specifically, we target at investigating novel vision transformers and generative deep learning models in order to derive molecular assays from H&E slides both in supervised and weakly-supervised manner. Furthermore, we plan to study how these spatial resolved molecular profiles correlate with treatment outcomes using a graph formulation under a deep learning paradigm. We aspire that the outcomes of this project can have a positive effect on clinical oncology and the therefore life expectancy of patients.