Beyond the Embedding Bottleneck: Adaptive Retrieval-Augmented 3D CT Report Generation

TL;DR

This paper identifies a representational bottleneck in 3D CT report generation caused by limited discriminative information in embeddings, and proposes AdaRAG-CT, an adaptive retrieval-augmented framework that improves clinical report quality.

Contribution

It introduces AdaRAG-CT, a novel adaptive retrieval-augmented method that enhances 3D CT report generation by compensating for visual representation limitations.

Findings

Achieves state-of-the-art Clinical F1 score of 0.480 on CT-RATE benchmark.

Demonstrates that adaptive retrieval improves report quality over static methods.

Confirms both retrieval and generation components are essential for performance gains.

Abstract

Automated radiology report generation from 3D CT volumes often suffers from incomplete pathology coverage. We provide empirical evidence that this limitation stems from a representational bottleneck: contrastive 3D CT embeddings encode discriminative pathology signals, yet exhibit severe dimensional concentration, with as few as 2 effective dimensions out of 512. Corroborating this, scaling the language model yields no measurable improvement, suggesting that the bottleneck lies in the visual representation rather than the generator. This bottleneck limits both generation and retrieval; naive static retrieval fails to improve clinical efficacy and can even degrade performance. We propose \textbf{AdaRAG-CT}, an adaptive augmentation framework that compensates for this visual bottleneck by introducing supplementary textual information through controlled retrieval and selectively integrating it during generation. On the CT-RATE benchmark, AdaRAG-CT achieves state-of-the-art clinical efficacy, improving Clinical F1 from 0.420 (CT-Agent) to 0.480 (+6 points); ablation studies confirm that both the retrieval and generation components contribute to the improvement. Code is available at https://github.com/renjie-liang/Adaptive-RAG-for-3DCT-Report-Generation.