A Bidirectional Loss Approach to Imparting Order Sensitivity to Multi-Image Chest X-ray Encoders

Doowoong Choi, Hanbin Ko, Jung Hoon Lee, Chang Min Park

Keywords: Deep Learning, Temporal Image Analysis, CXR

TL;DR: Bidirectional loss to ensure temporal order sensitivity in chest X-ray analysis, significantly improving consistency in longitudinal disease progression predictions.

Abstract:

Longitudinal chest X-ray (CXR) analysis is a critical step in assessing disease progression, yet existing deep learning methods often fail to account for the inherent temporal directionality of serial images, yielding inconsistent predictions when their order is reversed. In this work, we propose a bidirectional loss framework that enforces order sensitivity in multi-image CXR encoders. Leveraging large language models (LLMs), we obtain fine-grained interval change labels—resolved, improved, stable, worsened, and new—by comparing prior and current radiology reports across five common thoracic findings. We then exploit the symmetric nature of these labels by reversing image order and inverting labels (e.g., improved - worsened) during training. Experiments on the MIMIC-CXR and CheXpert datasets show that our method surpasses baselines for most findings, effectively embedding order awareness while retaining a simple, efficient design.

Submission Number: 42