By: Marie Arrivat¹², Rémy Peyret¹, Elsa Angelini², Pietro Gori²
(1 Primaa, Paris, 2 LTCI, Télécom Paris, Institut Polytechnique de Paris)

Context

• Whole Slide Images (WSIs) classifiers are often limited by biological complexity, reflected in diagnosis disagreement among pathologists.
• Collecting the opinion of several experts is expensive, but setting the ground truth with an expert and gathering the second opinion of a non-expert is easier.
• Whole Slide Difficulty (WSD) can be inferred and used as a prior to train Multiple Instance Learning (MIL) models.

Dataset

Use case: Gleason grading for prostate cancer

Whole Slide Difficulty (WSD)

Methodology

Performance

Attention Maps

Conclusion

Modelling WSD based on the disagreement between an expert and a non-expert pathologist and leveraging it in the loss function improves both the diagnostic accuracy and the interpretability of MIL models, especially for critical Gleason Grades.