A Review of Longitudinal Radiology Report Generation: Dataset Composition, Methods, and Performance Evaluation

TL;DR

This paper surveys longitudinal Chest X-ray radiology report generation (LRRG), addressing how historical studies are collected, processed, and leveraged to produce temporally coherent reports. It highlights three dataset paradigms (image-only, report-only, and combined modalities), various view configurations, and different scales of prior data that shape model inputs and evaluation. The review details three core components—vision encoders, text decoders, and loss formulations—while outlining major challenges such as missing data, cross-time alignment, and effective fusion, and surveys modern solutions including contrastive learning, fusion modules, and auxiliary prompts or retrievals. Through ablation analyses from LRRG studies, the paper emphasizes the critical role of longitudinal information and principled architectural choices in improving clinical fidelity, and it outlines limitations (data scarcity, temporal gap handling, redundancy, registration, and evaluation metrics) along with concrete directions for future work. Overall, this work provides a comprehensive foundation for advancing LRRG by clarifying best practices, benchmarking considerations, and promising research directions for robust, temporally aware radiology report generation.

Abstract

Chest Xray imaging is a widely used diagnostic tool in modern medicine, and its high utilization creates substantial workloads for radiologists. To alleviate this burden, vision language models are increasingly applied to automate Chest Xray radiology report generation (CXRRRG), aiming for clinically accurate descriptions while reducing manual effort. Conventional approaches, however, typically rely on single images, failing to capture the longitudinal context necessary for producing clinically faithful comparison statements. Recently, growing attention has been directed toward incorporating longitudinal data into CXR RRG, enabling models to leverage historical studies in ways that mirror radiologists diagnostic workflows. Nevertheless, existing surveys primarily address single image CXRRRG and offer limited guidance for longitudinal settings, leaving researchers without a systematic framework for model design. To address this gap, this survey provides the first comprehensive review of longitudinal radiology report generation (LRRG). Specifically, we examine dataset construction strategies, report generation architectures alongside longitudinally tailored designs, and evaluation protocols encompassing both longitudinal specific measures and widely used benchmarks. We further summarize LRRG methods performance, alongside analyses of different ablation studies, which collectively highlight the critical role of longitudinal information and architectural design choices in improving model performance. Finally, we summarize five major limitations of current research and outline promising directions for future development, aiming to lay a foundation for advancing this emerging field.

Figures (16)

• Figure 1: Examples illustrating the difference between single-image and LRRG settings. Single-image RRG: Without prior studies, the model fails to provide accurate comparison statements with prior examinations. LRRG: Access to prior studies enables the model to generate comparison statements closely aligned with those in the ground-truth report.
• Figure 2: Dataset construction strategies for the LRRG task, which can be divided into three categories: (a) Modality Composition of Prior Studies: prior studies may contain only images, only reports, or both. (b) Viewpoint Configuration: studies may consider only frontal views, multiple views for the current study, or multiple views for both current and prior studies. (c) Longitudinal dataset scale – Number of prior studies per case: each case may include either a single prior study or multiple temporally ordered prior studies.
• Figure 3: Examples from the MIMIC-CXR dataset demonstrate varying numbers of prior comparisons of different reports. Case 1 illustrates a single-time-point comparison, where findings are referenced only against the most recent prior study. Case 2 illustrates a multi-time-point comparison, where findings are evaluated against multiple prior studies, including the most recent, recent, and earlier dated radiographs.
• Figure 4: Hierarchical taxonomy of core challenges and technical strategies for longitudinal medical report generation.
• Figure 5: Strategies for unifying input sequence length in the LRRG task: (a) Using actual prior data when available; (b) Applying zero-padding when prior data is missing; (c) Placeholder tokens to indicate data absence; (d) Duplicating current data as a proxy for the missing prior input. All strategies ensure the same input token sequence length, enabling consistent model input formatting.