The paper “NCDD: Nearest Centroid Distance Deficit for Out-Of-Distribution Detection in Gastrointestinal Vision” has been accepted at the 2025 Medical Image Understanding and Analysis Conference (MIUA)!
It introduces a powerful new method to detect when deep learning models face unfamiliar or emerging gastrointestinal diseases—key to preventing overconfident, unreliable predictions. The authors propose the Nearest-Centroid Distance Deficit (NCDD) score, a novel approach tailored to the unique challenges of GI images, where disease features often overlap. Tested on leading benchmarks like Kvasir2 and GastroVision, NCDD outperforms current methods, paving the way for safer, more dependable AI in clinical endoscopy.
