Project overview
EQUR-XSM investigates how medical-image AI can communicate both its predictions and their reliability. It combines uncertainty estimation, confidence calibration and visual explanations, with applications in classification and segmentation. Led by Dawood Al Chanti at GIPSA-lab, this ANR JCJC project runs from October 2025 to October 2029.
Scientific summary
A medical-image model can produce an incorrect prediction with high confidence. A visual explanation alone does not resolve this problem: highlighting a region does not establish that the explanation is reliable. EQUR-XSM addresses these connected challenges by studying uncertainty in predictions and in the saliency maps used to explain them.
The project builds on Evidential Deep Learning (EDL) to estimate uncertainty and improve confidence calibration. It also investigates saliency maps enriched with pixel-wise uncertainty, aiming to provide information about influential image regions and the reliability of their interpretation.
A further objective is to combine complementary imaging modalities, such as MRI and CT, while accounting for the uncertainty associated with each input. The programme extends from image classification to single-modality and multimodal segmentation, targeting applications involving tumours, skeletal muscle abnormalities and haematological diseases.
Planned evaluation includes open datasets and clinical collaboration with the university hospitals of Grenoble and Nantes. The intended contribution is a set of methods for better-calibrated predictions, uncertainty-aware explanations and more reliable multimodal analysis. These are research objectives; clinical benefit remains to be established.