Lightweight and interpretable deep learning for reagent-free hemoglobin quantification in blood clam Anadara granosa

The blood clam Anadara granosa is notable among commercial bivalves for its unusually high hemoglobin content, which provides a rich source of bioavailable heme iron and underlies its deep-red hemolymph coloration relevant to product quality. Conventional spectrophotometric quantification requires laboratory infrastructure and chemical reagents, limiting its use in high-throughput screening and field-based breeding programs. In this study, we developed a reagent-free convolutional neural network (Hb-CNN) to estimate hemoglobin concentration directly from digital hemolymph images. Hemoglobin concentration was strongly associated with CIELAB color parameters, providing a colorimetric basis for image-based prediction. The optimized Hb-CNN achieved a low prediction error (MAE = 0.88 g/L) while maintaining a compact architecture with fewer than 3 million parameters and approximately 0.5 GFLOPs, demonstrating a favorable balance between predictive accuracy and computational efficiency. SHAP analysis further revealed that model predictions were primarily driven by lightness (L*) and redness (a*), with their contributions consistent with concentration-dependent light attenuation expected from Beer–Lambert optical principles and the intrinsic optical properties of hemoglobin, thereby providing a physicochemically interpretable basis for model decisions. Independent validation showed strong agreement with spectrophotometric measurements (R² = 0.86). A graphical interface further supported rapid batch processing and hemoglobin estimation. Overall, the lightweight and interpretable Hb-CNN provides a practical, reagent-free framework for hemoglobin phenotyping, with potential applications in seafood quality assessment and selective breeding.