Skip to content

PriorRG: Prior-Guided Contrastive Pre-training and Coarse-to-Fine Decoding for Chest X-ray Report Generation

Conference: AAAI 2026 arXiv: 2508.05353
Code: GitHub
Area: Medical Imaging / Radiology Report Generation Keywords: Chest X-ray report generation, prior knowledge, contrastive pre-training, coarse-to-fine decoding, spatiotemporal fusion

TL;DR

PriorRG proposes a two-stage chest X-ray report generation framework that aligns clinical context with spatiotemporal visual features via prior-guided contrastive pre-training, then progressively integrates clinical context, disease progression, and multi-level visual cues through prior-aware coarse-to-fine decoding, achieving a 3.6% improvement in BLEU-4 and a 3.8% improvement in F1 on MIMIC-CXR.

Background & Motivation

Automated radiology report generation (RRG) aims to reduce radiologists' workload by enabling AI systems to interpret medical images and produce structured textual descriptions. Most existing methods generate reports from a single image (see Figure 1(a) in the paper), overlooking patient-specific prior knowledge that radiologists routinely rely on in clinical practice, including:

Clinical Context (CC): indications and medical history, reflecting the diagnostic intent for a patient

Prior Image (PI): the most recent prior scan used to track disease progression

Several prior works have attempted to incorporate partial prior information but exhibit notable limitations: - Methods such as SEI leverage indication information but ignore longitudinal data, making the generated reports prone to hallucinations when describing disease progression. - Methods such as HERGen introduce prior images to model temporal changes but neglect clinical context, lacking personalized reasoning capability.

This motivates the central research question: Can temporal visual changes and clinical context be jointly modeled to improve cross-modal alignment and report generation?

Method

Overall Architecture

PriorRG adopts a two-stage training pipeline (see Figure 2 in the paper) that mirrors real-world clinical workflows: - Visual encoder: RAD-DINO (frozen) - Text encoder: CXR-BERT (trainable) - Report generator: DistilGPT2 (trainable)

Inputs include the current image \(x_i^{cur}\), prior image \(x_i^{pri}\) (possibly absent), indication \(z_i\) (possibly absent), and medical history \(h_i\) (possibly absent).

Key Designs

1. Stage 1: Prior-Guided Contrastive Pre-training

Objective: Leverage clinical context to guide spatiotemporal feature extraction and enhance cross-modal alignment.

Visual feature extraction: After extracting features with RAD-DINO, learnable view-position embeddings are introduced and fused into the visual features to handle appearance variation across different projection views (e.g., AP/PA), yielding \(\boldsymbol{V} \in \mathbb{R}^{M \times s \times d}\).

Text feature extraction: CXR-BERT encodes text with special tokens [INDICATION], [HISTORY], and [FINDINGS] prepended to indications, medical history, and reports respectively, enabling type-aware unified encoding while gracefully handling missing fields.

Spatiotemporal Fusion Network (STF): A ViT-style cross-attention fusion module models disease progression between the current and prior images:

\[\boldsymbol{V}^{st}_{ca} = \text{LN}(\boldsymbol{V}^{cur} + \text{CA}(\text{LN}(\boldsymbol{V}^{cur}), \text{LN}(\boldsymbol{V}^{pri})))\]
\[\boldsymbol{V}^{st} = \text{LN}(\boldsymbol{V}^{st}_{ca} + \text{FFN}(\boldsymbol{V}^{st}_{ca}))\]

where CA denotes cross-attention. The STF uses 3 layers. When no prior image is available, the current image features are used directly.

Instance-level cross-modal alignment: Simulating the clinical diagnostic process, a Perceiver architecture progressively fuses clinical context with spatiotemporal features:

\[\boldsymbol{\bar{T}}^c = \text{Perceiver}(\boldsymbol{E}^{lat}, \boldsymbol{T}^c)\]
\[\boldsymbol{\bar{V}}^{st} = \text{Perceiver}(\boldsymbol{\bar{T}}^c, \boldsymbol{V}^{st})\]

Global visual features \(\boldsymbol{V}^g\) are obtained via global average pooling and L2 normalization. Image–report similarity is computed and optimized using a cross-entropy alignment loss \(\mathcal{L}_{align}\) that supports multiple positive pairs.

2. Stage 2: Prior-Aware Coarse-to-Fine Decoding

Attention-enhanced Layer Fusion Network (ALF): CBAM-based channel and spatial attention is applied to features from each encoder layer to highlight diagnostically relevant information. A Conv2D projector then fuses these into a multi-level visual representation \(\boldsymbol{V}^{hier}\), addressing the limitation of existing methods that rely solely on the last hidden state and overlook low-level details such as lesion morphology.

Coarse-to-fine decoding: Inspired by principles of visual cognition, prior knowledge and multi-level visual semantics are progressively integrated: - Coarse-grained priors: \(\boldsymbol{\bar{T}}^c\) (clinical context) and \(\boldsymbol{\bar{V}}^{st}\) (spatiotemporal features) supply high-level clinical background and disease progression cues. - Fine-grained enhancement: \(\boldsymbol{\bar{V}}^{hier} = \text{Perceiver}(\boldsymbol{\bar{V}}^{st}, \boldsymbol{V}^{hier})\), where spatiotemporal features serve as queries to extract fine-grained information from multi-level features. - Final concatenation: \(\boldsymbol{\bar{T}}^c\), \(\boldsymbol{\bar{V}}^{st}\), and \(\boldsymbol{\bar{V}}^{hier}\) are concatenated along the sequence dimension and fed into DistilGPT2 for report generation.

Loss & Training

  • Stage 1: Contrastive alignment loss \(\mathcal{L}_{align}\) (cross-entropy), supporting multi-view positive pairs.
  • Stage 2: Cross-entropy loss \(\mathcal{L}_{CE}\) for training autoregressive report generation.
  • Unified feature dimension \(d=768\), number of latent variables \(N=128\), maximum generation length \(K=100\), beam size 3.
  • AdamW optimizer with ReduceLROnPlateau scheduler and early stopping (patience = 15).
  • On MIMIC-CXR: 30 epochs for Stage 1 pre-training and 30 epochs for Stage 2 fine-tuning.

Key Experimental Results

Main Results

Dataset Metric PriorRG Prev. SOTA Gain
MIMIC-CXR B-1 0.412 0.416 (MPO) -0.4%
MIMIC-CXR B-4 0.175 0.139 (MPO) +3.6%
MIMIC-CXR MTR 0.189 0.176 (BioViL-T) +1.3%
MIMIC-CXR R-L 0.324 0.309 (MPO) +1.5%
MIMIC-CXR F1 0.511 0.473 (R2-LLM) +3.8%
MIMIC-ABN B-1 0.326 0.267 (SEI) +5.9%
MIMIC-ABN B-4 0.102 0.073 (SEI) +2.9%
MIMIC-ABN F1 0.471 0.460 (CMN) +1.1%

PriorRG achieves comprehensive superiority on long n-gram matching (B-4) and clinical accuracy (F1), outperforming SEI on 13 of 14 CheXpert observation categories in terms of F1.

Ablation Study

Configuration B-4 F1 Notes
(a) No CC, No PI, No Hidden 0.108 0.472 Baseline: Stage 1 only, no prior knowledge
(c) CC, No PI, No Hidden 0.170 0.487 Adding clinical context yields significant gains
(e) CC, PI, No Hidden 0.171 0.499 Prior image provides further improvement
(d) CC, No PI, Hidden 0.173 0.507 Multi-level features improve clinical accuracy
(f) No Stage 1, Stage 2 only 0.165 0.459 Omitting pre-training causes substantial degradation
PriorRG (full) 0.175 0.511 All components synergistically optimal

Key Findings

  1. Clinical context has the largest impact: As CC availability increases from 0% to 100%, B-2 rises from 0.139 to 0.294 (Table 5), demonstrating that indications and medical history are critical for report generation.
  2. Prior image contribution is more pronounced at the study level: It enhances temporal cue modeling, yielding significant improvements on the Stu-P@K retrieval metric.
  3. Coarse-to-fine outperforms fine-to-coarse: PriorRG achieves better NLG metrics than the Fine2coarse variant, confirming that progressively integrating high-level semantics before fine-grained details is more effective.
  4. Zero-shot report generation capability: In an unsupervised setting, PriorRG achieves B-4 = 0.178 and MTR = 0.211, substantially outperforming R2GenGPT and Med-LLM.

Highlights & Insights

  1. Complete simulation of clinical workflow: The pipeline — from initial clinical assessment to spatiotemporal contrast to multi-level refinement — closely mirrors radiologists' diagnostic reasoning process.
  2. Graceful handling of missing inputs: Special tokens and the Perceiver architecture naturally accommodate the absence of prior images, indications, or medical history, making the framework practically deployable.
  3. Disease progression description capability: Qualitative analysis shows that PriorRG correctly describes lesion changes (e.g., "cardiac silhouette remains enlarged but unchanged"), reducing hallucinations when characterizing temporal variations.
  4. Leading GREEN scores: Evaluated using the pretrained GREEN-RadLlama2-7B model, PriorRG significantly outperforms all baselines on both the number of matched findings and the overall GREEN composite score.

Limitations & Future Work

  1. The use of DistilGPT2 as the report generator constrains capacity due to its relatively small model scale; future work may explore more powerful LLMs such as LLaMA.
  2. Validation is limited to the MIMIC-CXR dataset, with no exploration of report generation for other imaging modalities (e.g., CT, MRI).
  3. Organ-level localization information is not incorporated; the authors note in the conclusion that an organ-aware diagnostic framework will be explored in future work.
  4. The two-stage training process is sequential; end-to-end joint optimization may yield further performance improvements.
  • BioViL-T / MLRG: Pioneering works in longitudinal data modeling, but lacking clinical context.
  • SEI: A representative method utilizing indication information, but neglecting disease progression.
  • Perceiver architecture: Effectively employed for progressive cross-modal fusion, serving as a versatile tool for connecting information at different levels of granularity.
  • Insight: The impact of clinical prior knowledge — especially indications — on report generation far exceeds expectations; visual information alone is insufficient.

Rating

Dimension Score (1–5) Notes
Novelty 4 First work to fully exploit the combination of clinical context, prior images, and multi-level visual features
Technical Depth 4 Two-stage design is well-motivated; Perceiver-based progressive fusion is creative
Experimental Thoroughness 5 Covers main experiments, ablations, retrieval, qualitative analysis, and LLM-based evaluation
Value 4 Directly aligned with clinical workflows; handles missing inputs effectively
Writing Quality 4 Well-structured, richly illustrated, with clearly articulated motivation