PhD defense Ali Mammadov: Role and Robustness of Self-Supervised and Multiple Instance Learning Approaches for Digital Pathology
Télécom Paris, 19 place Marguerite Perey F-91120 Palaiseau [getting there], amphi 2 and in videoconferencing
Full title: On the Role and Robustness of Self-Supervised and Multiple Instance Learning Approaches for Digital Pathology – Application to Sjogren’s Syndrome
Jury
-
Thomas Walter, Professor, École des Mines de Paris (Reviewer)
-
Nicolas Loménie, Professor, Université Paris Cité (Reviewer)
-
Daniel Racoceanu, Professor, Sorbonne University (Examiner)
-
Maria Vakalopoulou, Assistant Professor, CentraleSupélec (Examiner)
-
Daniel Stockholm, Associate Professor, École Pratique des Hautes Études (Examiner)
-
Pietro Gori, Professor, Télécom Paris (Thesis Director)
-
Loïc Le Folgoc, Associate Professor, Télécom Paris (Thesis Co-Supervisor)
Abstract
Digital pathology allows tissue samples to be scanned into Whole Slide Images (WSIs), enabling more detailed and accurate medical diagnosis. These images are extremely large, making manual analysis by pathologists time-consuming and challenging. Artificial intelligence (AI) can help by automatically examining these images to detect diseases efficiently and reliably. This study explores Multiple Instance Learning (MIL), a method that treats each slide as a collection of smaller image patches.
Traditionally, MIL methods either classify each patch independently or combine all patches into a single representation. While the independent approach is more interpretable, the combined approach has often been preferred due to performance limitations. With advances in new self-learning AI methods, patch-based methods can now achieve similar or better results while remaining easier to interpret. Another challenge is variability in AI model performance caused by random training factors. A Multi-Fidelity Model Fusion strategy was developed to combine the most stable models, improving reliability and reproducibility. These methods were applied to Sjogren’s syndrome, an autoimmune disease affecting saliva and tear glands. A fully automated pipeline detects relevant cell clusters in tissue samples and calculates a focus score, replicating the pathologist’s workflow. This approach provides accurate, interpretable, and clinically aligned diagnostics, demonstrating the potential of AI to support faster and more reliable medical decisions.