Proof-of-concept for a deep learning-based automated system for deep-sea fish stomach content analysis

In fish dietary analysis, reducing the substantial time and effort required is essential for accelerating the accumulation of ecological knowledge. Recent advances in machine learning have enabled development of highly accurate object detection models. Here, we aimed to establish a foundational technology for automating stomach content analysis by developing an automated prey detection system based on YOLO, a general-purpose object detection model. Stomach contents from demersal fish collected by bottom trawling were photographed using a stereomicroscope, and a training dataset comprising 682 images (1655 instances) across 18 classes was constructed. The developed model successfully detected prey items across a wide range of digestion states, including highly degraded and amorphous prey that are difficult to identify visually. Despite the challenging nature of this task, the class-average precision, recall, and class assignment error rate were 41.3% ± 8.4%, 37.2% ± 7.6%, and 34.1% ± 7.2%, respectively. Among all classes, precision was highest for the ‘Fish(eye)’ class at 88.3%, whereas recall was highest for ‘Decapoda(shrimp)’ at 70.5%. In addition to these conventional metrics, we introduced a novel metric, Identification to Higher-level Taxon (IdHT), to account for taxonomic-level inconsistencies common in dietary analyses. The class-average IdHT was 19.5% ± 5.3%. Inter-class confusion occurred frequently, particularly among crustacean classes, likely due to texture bias inherent in convolutional neural networks. This study represents the first application of deep learning–based object detection to dietary analysis across multiple prey taxa, providing a foundational framework for efficiently conducting large-scale deep-sea food-web studies in the future.