Close Mobile Navigation


Ce projet est porté par le Centre Léon Bérard et la société OWKIN


DIAGNOSE is a research project aiming to establish novel AI-based tools.

The diagnosis of rare cancers is challenging, cause of misdiagnosis and inappropriate treatment and reduced survival. The nosological classification of human cancer is based on a combination of histological analysis, immunohistochemistry and the presence of genomic alterations of the cancer cell. Cancer classification has been complexified by the characteristic of the tumor microenvironment important predictors of sensitivity to immune checkpoint treatments. The aims of D-IA-Gnose are: to improve diagnostic accuracy in rare cancer, to decipher complexes histologies, to identify new markers predictive of outcome, to optimize prediction of the tumor microenvironment and treatment response in two rare cancer models where the CLB has large expertise: sarcomas and mesotheliomas. Since 2017, a partnership was established between  Owkin and the Centre Léon Bérard. The rapid implementation of AI-based tools using H&E, a panel of Antibodies, Sequencing (WES and WTS) is now moving to the clinic for a more accurate classification improving the current standards of clinical practice.


SGC & GIST : Prediction of prognosis in SGC (complex genomic sarcoma).

The current prognosis tool used in clinical practice to evaluate the risk of metastasis relapses for Soft Tissue Sarcoma patients is a nomogram called Sarculator. It is based on the patient’s age, the tumor grade, the tumor size and the histological evaluation. The primary aim is to use Artificial Intelligence on Whole Slide Images (WSI) combined with clinical data to develop a prognosis score that outperforms the Sarculator. Clinically, this tool would be able to improve the indication of chemotherapy. The secondary goal is to assess the benefits of a federated learning setting for training deep learning models in separate centers.


Using deep learning methods on histology, models were built on WSI to predict Overall Survival and Metastasis-Free Survival and in local settings. They outperform the Sarculator nomogram. A single model is now developing in a federated learning setting to validate its performance in a multicenter approach and strengthen its robustness for future use.



The first step was to evaluate outcomes indicators using Deep learning on H&E of WSI (Whole Slide Images) from the MESOBANK/MESOPATH published in Nature Medicine in 2019.


Then, to further evaluate the potential of deep learning at the molecular level, we analyzed WTS (Whole Transcriptomic Sequencing) from 198 patients retrieved from Franck Tirode’s database. We realized a supervised approach and confirmed the performance of MESONET prediction on this new dataset of WSIs that were similar to pathologists classification (C-index MESONET = 0.60 versus C-index Histomean = 0.62). Then to gain insights into the molecular mechanisms underlying histological heterogeneity, the consensus hierarchical clustering of all tumors using the same RNAseq data were able to individualize 6 clusters instead of the three histological subtypes, based on their molecular signatures epithelioid and sarcomatoid scores with one cluster (CL3 being immune cold). Later we used Microenvironment Cell Population Counter (MCP-counter) to assess the proportion of several immune cell populations in the tumor microenvironment. This analysis demonstrated that tumors from high-risk MESONET patients were significantly associated with lower B cell infiltrations than their low-risk counterparts. As a result, Differentially expressed (DE) genes analyses identify genes overexpressed among CL3-tumors versus non- CL3-tumors for future identification of innovative therapy targets.


1. Blay JY. Editorial: Customizing sarcoma management and treatment. Curr Opin Oncol. 2019 Jul;31(4):302-303.

2. Courtiol P, Maussion C, Moarii M, Pronier E, Pilcer S, Sefta M, Manceron P, Toldo S, Zaslavskiy M, Le Stang N, Girard N, Elemento O, Nicholson AG, Blay JY, Galateau-Sallé F, Wainrib G, Clozel T. Deep learning-based classification of mesothelioma improves prediction of patient outcome. Nat Med. 2019 Oct;25(10):1519-1525.

3. Galateau Salle F, Le Stang N, Tirode F, et al, Comprehensive Molecular and Pathologic Evaluation of Transitional Mesothelioma. JTO 2020 Jun;15(6):1037-1053.