A multi-stage deep learning framework for automated detection and localization of Shrimp disease

Abstract

This work presents an extensive Multi-Stage Deep Learning Framework to reduce the significant financial effects of White Spot Syndrome Virus (WSSV) and Black Gill (BG) disease on the $64 billion global shrimp aquaculture sector by fixing the interpretive problems and restricted ability to diagnose ability of current automated systems. The study develops an ordered detection pipeline starting with disease detection using a dataset of 5,498 high-resolution images. Following a comparison of five state-of-the-art architectures, ConvNeXt-Tiny proved to be the best classifier. It accomplished a test accuracy of 96.8% and a Macro F1-score of 0.97. To solve baseline weaknesses, Projected Gradient Descent (PGD) adversarial training was executed to restore robust accuracy to 55.5% against environmental disturbances. After classification of disease, a conditional disease specific U-Net++ segmentation module had been developed for precise lesion identification, giving a Dice Score of 0.8217 for Black Gill and 0.71 for WSSV, which enabled the development of a novel pixel-level Severity Staging system (Mild, Moderate, Severe) for actionable intervention. This research successfully bridges the gap between theoretical computation and dependable, practical application by combining Explainable AI (XAI) with Grad-CAM visualisations for assessing histological features and optimizing for edge-deployment efficiency. This has been achieved by both of these techniques.