Deep-learning histopathradiomics enables high-throughput quantification of pollutant-induced tissue injury in marine mussels.

Quantitative tissue-level biomarkers are important for assessing pollutant-induced biological effects, yet conventional histopathology remains largely manual, low-throughput, and observer-dependent. We developed a deep learning-based histopathradiomics workflow for high-throughput morphometric quantification in the marine mussel Mytilus galloprovincialis. Benzo[a]pyrene (BaP) was used as a benchmark stressor to provide a biologically defined tissue-injury context for workflow evaluation. An optimized Res-UNet achieved Dice coefficients of 0.9159 ± 0.0788 for gill epithelium and 0.8687 ± 0.1612 for digestive tubule walls, enabling automated quantification of 1,264 gill filaments and 1,437 digestive tubules within minutes. The resulting gill epithelial area and digestive tubule wall area showed injury-related patterns consistent with semi-quantitative pathology and manual thickness-based morphometry. The current binary framework focuses on predefined structural endpoints rather than the full spectrum of histopathological lesions. Without further model optimization, application to an independent sulfamethoxazole (SMX)+BaP dataset retained comparable tissue-injury patterns across exposure scenarios and independent experimental and histological batches, supporting cross-scenario applicability. Overall, this workflow transforms established histopathological biomarkers into scalable and biologically interpretable morphometric endpoints, while reducing repeated manual delineation and providing a practical basis for standardized tissue-level effect assessment in marine sentinel organisms.