Deep learning and hyperspectral imaging for non-destructive amino acid detection in live carp fillets

Rapid and non-destructive inspection of fillet nutritional quality is essential for selecting high-value live fish prior to processing, yet no such method exists for determining fillet amino acid (AA) contents. This study developed a non-destructive approach integrating visible/near-infrared (Vis/NIR) hyperspectral imaging of fish scales with deep learning (DL) models to predict fillet AA contents in live common carp (Cyprinus carpio). Scale spectra (400-1000 nm) and corresponding dorsal muscle AA profiles were collected from 481 fish across two distinct populations. Five DL models were compared, with a backpropagation artificial neural network (BP-ANN) demonstrating superior performance. While sample population heterogeneity was the primary factor reducing model accuracy (mean R² decrease of 0.182), the BP-ANN models remained robust, achieving validation R² values exceeding 0.777 for all AAs. For the three most abundant AAs (glutamic acid, aspartic acid, and lysine), the BP-ANN model achieved validation R² values of 0.848, 0.858, and 0.858, respectively. Competitive adaptive reweighted sampling (CARS) identified characteristic wavelengths for glutamic acid and lysine primarily in the 516-584 nm, 707-738 nm, 828-834 nm, and 939-1032 nm regions. Our approach revealed uneven distribution across the fish body, with mandibular, pectoral, and abdominal muscles exhibiting the highest total AA (TAA) content. Notably, fillet TAA levels remained stable across fish of different sizes despite variations in dietary protein content. This rapid, non-destructive, and cost-effective approach enables real-time nutritional grading and sorting of live fish, offering a practical tool for precision aquaculture and high-quality live fish selection in food industry.