A Clinically Aligned Local-global Mutual Learning Framework for Interpretable Pneumoconiosis Diagnosis

Pneumoconiosis is a life-threatening occupational lung disease requiring accurate and timely diagnosis. However, the clinical interpretation of chest radiographs remains challenging due to subtle lesion patterns and substantial inter-observer variability. While deep learning methods have shown promise, they often lack interpretability and fail to incorporate clinical diagnostic knowledge, making their predictions difficult to trust and integrate into clinical workflows. Additionally, the limited scale and diversity of available datasets restrict the generalizability of these models. To address these challenges, we propose a novel clinically aligned local-global mutual learning (CALGM) framework tailored for interpretable pneumoconiosis diagnosis. CALGM integrates a clinically guided hierarchical classification mechanism that mimics clinical diagnostic reasoning by decomposing the global image assessment into zone-level severity evaluations, followed by rule-based local-global ensembling inference. The framework further incorporates diagnostic constraints that regularize attention toward clinically relevant lung regions, enforce ordinal consistency in severity progression across zones, and align local zone-level attention with the global lung context. These designs collectively improve model discriminability, reduce false positives, and ensure decision transparency in alignment with clinical protocols. Extensive experiments on internal and external pneumoconiosis datasets, as well as the public the Japanese society of radiological technology (JSRT) dataset, demonstrate that CALGM outperforms state-of-the-art methods across key performance metrics. These results highlight CALGM′s strong generalization ability and its potential for reliable, interpretable deployment in clinical practice. The code is available at https://github.com/likakakaka/CALGM.