TRACE: Temporal Radiology with Anatomical Change Explanation for Grounded X-ray Report Generation

TL;DR

This paper addresses the need for temporal reasoning in chest X-ray analysis by introducing TRACE, a model that jointly performs temporal change detection, change classification, and spatial grounding between prior and current studies. TRACE encodes both images with a frozen BioViL-T encoder, fuses their features, and uses a LoRA-tuned Vicuna-7B language model to generate grounded interval-change descriptions. A large temporal grounding dataset is constructed from MIMIC-CXR-JPG and Chest ImaGenome, enabling evaluation of change detection, grounding, natural language generation, and clinical validity. The study reveals an emergent property: change detection only arises when both temporal input and grounding supervision are present, and demonstrates strong grounding (IoU>0.5 mean 0.772) and competitive change-detection performance (48.0% accuracy), along with meaningful clinical reporting.

Abstract

Temporal comparison of chest X-rays is fundamental to clinical radiology, enabling detection of disease progression, treatment response, and new findings. While vision-language models have advanced single-image report generation and visual grounding, no existing method combines these capabilities for temporal change detection. We introduce Temporal Radiology with Anatomical Change Explanation (TRACE), the first model that jointly performs temporal comparison, change classification, and spatial localization. Given a prior and current chest X-ray, TRACE generates natural language descriptions of interval changes (worsened, improved, stable) while grounding each finding with bounding box coordinates. TRACE demonstrates effective spatial localization with over 90% grounding accuracy, establishing a foundation for this challenging new task. Our ablation study uncovers an emergent capability: change detection arises only when temporal comparison and spatial grounding are jointly learned, as neither alone enables meaningful change detection. This finding suggests that grounding provides a spatial attention mechanism essential for temporal reasoning.

Figures (3)

• Figure 1: Comparison of chest X-ray report generation approaches. Single-Image Report Generation (R2Gen chen2020generating) generates reports without temporal context or spatial grounding. Temporal Comparison (BioViL-T bannur2023learning, CheXRelNet karwande2022chexrelnet, CheXRelFormer mbakwe2023chexrelformer) detects disease progression but outputs only classification labels. Grounded Report Generation (ChEX muller2024chex) produces spatially grounded reports but lacks temporal reasoning. Temporal Grounded Report Generation (MAIRA-2 bannur2024maira2) uses prior images with grounding but without explicit change classification. TRACE (Ours) is the first to combine explicit temporal change classification with spatial grounding, describing what changed, where, and how.
• Figure 2: Overview of the TRACE architecture. Prior and current chest X-rays are encoded separately by a frozen BioViL-T encoder with shared weights. The resulting feature sequences (196 tokens each) are concatenated to form 392 visual tokens, projected via a trainable MLP, and decoded by a large language model with LoRA fine-tuning to generate grounded temporal reports.
• Figure 3: Qualitative results of TRACE on three representative cases: worsening, stable, and improvement. Each row shows the prior image, current image, model prediction with bounding box overlay, and the corresponding ground truth and model output text. The model successfully identifies temporal changes and localizes the affected anatomical regions with accurate bounding box coordinates.