Deep learning driven smartphone system for fish disease diagnosis in Indian major carps

Abstract

Fish disease poses a substantial threat to the productivity and sustainability of Indian major carps (IMCs) in aquaculture. Early and accurate fish disease detection is crucial to minimizing economic losses and ensuring fish health management. In this study, we developed a lightweight deep learning model, FishMoViT, for real-time detection of seven major IMCs diseases. It integrates two state-of-the-art deep learning networks, MobileNetV2 and MobileViT-Small, as a dual backbone, leveraging an ensemble strategy to combine complementary feature representations. The dataset comprising seven major disease classes was randomly split into training, validation, and test sets at 70%, 10%, and 20%, respectively, for the model development. The dataset comprises 1,252 images and expanded to 6,272 images through data augmentation techniques after splitting. The proposed FishMoViT model achieved superior performance with an average precision, recall, accuracy, and F 1 -score of 96% each. We observed near-perfect classification for Argulosis, Septicemia, Fin and tail rot, and Cotton wool, while relatively higher misclassification for Gill rot and Lernaeosis due to visual similarities. The ablation study shows that integrating both backbone networks improved classification accuracy by 5.14% and 2.09% compared with the MobileNetV2-only and MobileViT-Small-only variants, respectively. Compared with MobileNetV1, MobileNetV3-Large, InceptionV3, and VGG-16, FishMoViT improved classification accuracy by 9.01%, 12.55%, 8.93%, and 11.98%, respectively, while maintaining a lightweight model size of only 27.65 MB. Statistical validation using McNemar’s test confirmed that FishMoViT significantly outperformed all baseline models (χ² = 53.79–239.02, p  < 0.001). The comparison with state-of-the-art vision transformer models (DeiT, SWIN) further validates the architectural choice of FishMoViT. To demonstrate practical applicability, we developed an Android-based mobile application, AquaScan, that deploys the FishMoViT model for real-time fish disease diagnosis and management. The findings demonstrate that the proposed lightweight FishMoViT model and AquaScan enable precise, efficient, and field-ready fish disease detection, making it a viable approach for supporting sustainable aquaculture management.