We are delighted to announce that 1 paper from VI Lab (Prof. Jongwon Choi) has been accepted to NeurIPS 2026.

Title: 

FinAl: Fine-grained Alignment for Detail-Preserving Medical Vision-Language Pretraining

Authors: 

Jongsu Youn, Dongyoung Lee, JaeHyeong Bae, Sang-Il Choi*, Sangtae Choi*, Dae Ung Jo*, Jongwon Choi*

Abstract: 

Medical vision-language pretraining with chest X-rays and radiology reports can overlook fine clinical distinctions when all unmatched image-text pairs are treated as equally negative. FinAl uses relative semantic relationships among structured radiology reports to preserve graded similarities, including differences in disease severity. It further adjusts alignment strength as the visual and language representations evolve, without requiring explicit severity labels during pretraining. Reported chest X-ray severity-classification and image-text retrieval evaluations show improvements in fine-grained clinical understanding over existing approaches.