A Prompt-Guided Vision-Language Framework for Interpretable and Region-Aware Disease Diagnosis in Chest X-rays.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Current machine learning systems for chest X-ray interpretation often process visual analysis, diagnostic reasoning, and reporting in isolation.
• This separation leads to a lack of diagnostic context in visual encoders and spatial grounding in language outputs.

Purpose of the Study:

• To develop an interactive vision-language framework for region-aware and clinically aligned chest X-ray interpretation.
• To enable prompt-guided reasoning over both textual and spatial queries for enhanced diagnostic support.

Main Methods:

• The proposed framework integrates three modules: Prompt-Guided Localization (PGL), Region-Level Diagnosis (RLD), and Region-Aware Explanation (RAE).
• A multi-task Detection Transformer (DETR) backbone unifies these modules via a regional alignment mechanism, mapping prompts and image regions into a shared semantic space.
• A two-stage training strategy involving contrastive pretraining and multi-task fine-tuning was employed for limited supervision.

Main Results:

• The framework demonstrated consistent performance gains across multiple public chest X-ray datasets (MIMIC-CXR, VinDr-CXR, MS-CXR) compared to state-of-the-art methods.
• Ablation studies confirmed the significant contribution of each module to the overall performance.
• The system achieved clinically aligned interpretations and improved disease classification and report generation.

Conclusions:

• The interactive vision-language framework effectively integrates visual and textual information for chest X-ray interpretation.
• The proposed approach offers potential for transparent and clinically applicable diagnostic support in radiology.
• Future work can further explore the integration of spatial and textual reasoning in medical AI.