ECHO: Efficient Chest X-ray Report Generation with One-step Block Diffusion
Source: arXiv:2604.09450v1 | Type: Cross
Abstract
Chest X-ray report generation (CXR-RG) has the potential to substantially alleviate radiologists’ workload. However, conventional autoregressive vision-language models (VLMs) suffer from high inference latency due to sequential token decoding. Diffusion-based models offer a promising alternative through parallel generation, but they still require multiple denoising iterations. Compressing multi-step denoising to a single step could further reduce latency, but often degrades textual coherence due to the mean-field bias introduced by token-factorized denoisers.
Introduction
In recent years, the demand for efficient medical imaging analysis has surged, highlighting the need for advanced artificial intelligence (AI) solutions. Traditional methods of generating reports from chest X-rays are not only time-consuming but also require extensive manual intervention. The introduction of ECHO represents a significant milestone in this domain, offering an innovative approach to streamline the report generation process.
The ECHO Framework
We propose ECHO, an efficient diffusion-based vision-language model (dVLM) specifically designed for chest X-ray report generation. ECHO enables stable one-step-per-block inference through a novel framework known as Direct Conditional Distillation (DCD). This framework addresses the challenges associated with mean-field limitations by constructing unfactorized supervision from on-policy diffusion trajectories, effectively encoding joint token dependencies.
Key Features of ECHO
- One-Step Block Inference: ECHO achieves significant speed improvements by allowing for one-step inference for each block, reducing the overall processing time.
- Direct Conditional Distillation: The DCD framework mitigates the mean-field bias, facilitating better generation of coherent reports.
- Response-Asymmetric Diffusion (RAD): This training strategy enhances efficiency while preserving the effectiveness of the model, leading to more accurate report generation.
Performance Evaluation
Extensive experiments have demonstrated that ECHO outperforms state-of-the-art autoregressive methods. Key performance metrics include:
- RaTE Improvement: ECHO improves the Report Accuracy through Tokens Efficiency (RaTE) by 64.33%.
- SemScore Enhancement: The Semantic Score (SemScore) sees an improvement of 60.58%.
- Inference Speed: ECHO achieves an incredible 8× speedup in inference time without sacrificing clinical accuracy.
Conclusion
The introduction of ECHO marks a significant advancement in chest X-ray report generation technology. By leveraging innovative diffusion-based techniques and enhancing the speed and accuracy of report generation, ECHO demonstrates the potential of AI to alleviate the burden on radiologists. As AI continues to evolve, solutions like ECHO pave the way for more efficient and effective healthcare delivery.