Title: Unified Multi-Task Framework Facilitates Interpretable Chest Radiograph Evaluation

Abstract: Although multimodal deep learning has significantly propelled the field of medical imaging analysis, current black-box models are frequently restricted to singular tasks. This limitation often neglects the multi-task nature of clinical diagnosis, which demands high levels of trust. To address this, we introduce IMT-CXR (Interpretable Multi-task Transformer for Chest X-ray Analysis), a novel framework that mirrors the diagnostic workflow of radiologists. This system operates through three evidence-based phases: 1) identifying diseases; 2) characterizing attributes, such as quantifying severity, location, and size; and 3) generating reports that integrate evidence with traceable decision pathways.

Built on a unified transformer architecture and refined through medical-domain instruction tuning, the framework sequentially performs four specific clinical tasks: radiology report generation, anatomical segmentation, lesion localization, and multi-label disease classification. Our experimental results indicate competitive performance across ten CXR benchmarks in both fine-tuning and direct inference scenarios. Furthermore, a blinded assessment of 160 historical reports sourced from four medical centers revealed that three radiologists judged 66% of the AI-generated reports to be equal to or better than the original clinical reports in terms of diagnostic clarity. This finding underscores the framework’s potential for clinical translation. By creating traceable diagnostic paths that connect anatomical observations to final conclusions, this research narrows the divide between AI technical metrics and clinical utility, thereby promoting the development of trustworthy AI systems in medical imaging.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC