Physician-Guided Attention Refinement in Chest Computed Tomography Classification under Label Ambiguity

Yueh-Chun Liu, Chia-Jung Liu, Yu-Hsuan Chen, Chang-Wei Wu, Meng-Rui Lee, Po-Chih Kuo

Keywords: Computed Tomography, Human-in-the-Loop, Expert-Guided Learning, Attention Mechanism

TL;DR: We propose a physician-guided human-in-the-loop framework that refines model predictions and attention through expert feedback, improving generalization and diagnostic performance under label ambiguity in chest CT.

Registration Requirement: Yes

Abstract: Annotation ambiguity due to inter-physician variability remains a major challenge in medical image analysis. In diseases with ambiguous imaging patterns, diagnosis often requires multi-physician consensus, leading to soft and inconsistent labels. We propose a physician-guided human-in-the-loop (HITL) framework that iteratively refines model predictions through expert feedback. Our approach integrates a 3D convolutional neural network (CNN) with an attention mechanism, where physicians review predictions and Grad-CAM attention heatmaps to provide reassessments and feedback for iterative refinement. Experiments on a private multi-center chest computed tomography (CT) dataset of patients with non-tuberculous mycobacterial (NTM) infection demonstrate improved performance, with internal sensitivity increasing from 0.71 to 0.82 under comparable AUROC, and external AUROC improving from 0.72 to 0.80 alongside sensitivity from 0.66 to 0.80. Physician consensus also improves by over 35\% in ambiguous cases. These results highlight the value of incorporating structured expert feedback to improve model performance and clinical decision support in uncertain diagnostic scenarios.

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 82