An Edge-Optimized Deep Learning Framework for Real-Time Aquatic Pathology Detection and Automated Therapeutic Guidance

The quick and precise detection of aquatic diseases is an essential component to support the globally thriving aquaculture sector, while at the same time providing security for aquarium inhabitants. Currently, the veterinary diagnosis of fish disease is performed manually, is quite subjective and not always available, especially in remote or poor environments. The last few applications have focused on the classification of fish disease using deep learning, but it relies heavily on time consuming cloud-dependent networks and is only able to classify the disease with a predicted class as the result. This reseach proposes AquaDiagnost, a highly optimized and end-to-end deep learning framework with built-in automatic treatment expert, focusing on edge devices. The technique show that with a novel two-stage transfer learning methodology trained over MobileNetV2 network, the framework can achieve up to 7 fresh-water fish diseases identification with only visual symptoms being analyzed, the weighted F1-score and accuracy is 0.90 and 90.2% with the dataset of 2,444 images respectively. A compact system with only 3.5M parameters, it achieves less than 0.8 seconds inference per image on the local devices. Moreover, prediction result from the proposed framework automatically invokes a diagnostic decision rule engine, and then immediate generate deterministic treatment protocol including specific dosages of medication, first step actions for quarantining affected fish, changes in physical and chemical water parameters. It turns out that AquaDiagnost presents an innovative approach for low-latency, privacyprotecting early detection of aquatic disease at the edge.