Clinical Attention Consistency for Cross-Hospital Robustness in Vision Transformers for Multi-Label Chest X-Ray Classification
Year: 2026
Type: preprint
Abstract: Chest X-ray analysis models trained at one medical institution frequently fail to generalize to others due to domain shift. We propose a Clinical Attention Consistency (CAC) loss that supervises Vision Transformer (ViT-B/16) attention maps using radiologist bounding box annotations, encouraging the model to focus on clinically relevant anatomical regions. We train on the NIH Chest X-Ray dataset (112,120 images, 14 disease labels) and evaluate cross-hospital generalization on CheXpert. Our baseline ViT-B/16 achieves AUC 0.7664 on the NIH test set. Cross-hospital evaluation shows improvement in 5 of 7 shared disease categories. We identify annotation sparsity as a key limitation and propose weighted random sampling as a concrete solution.
Author: Ehtasham, Arif
Language: English
Topic: COVID-19 diagnosis using AI
Subfield: Radiology, Nuclear Medicine and Imaging
Field: Medicine
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W7156089359