Red mark syndrome (RMS) is a skin condition characterized by deep, chronic dermatitis associated with panniculitis and myositis, affecting mainly rainbow trout (Oncorhynchus mykiss ) significantly impacting the aquaculture industry . The aquatic environment where fish live presents challenges for diagnosing and treating water-borne diseases. Traditional diagnostic methods are often time-consuming and used when the disease has progressed significantly. Therefore, early and accurate diagnosis is crucial to minimize the impact of the disease on fish populations. In this study, we tested the possibility of early detection of disease symptoms by camera technology under real conditions. RMS was used since the pathology mainly manifests itself externally and in addition is fairly distinct from pathology with other causes.
In total, 160 r ainbow trouts (Oncorhynchus mykiss) were used in our study. The size of the fish at the cohabitation with the seeders was 22.5 ± 1.4 cm 150 ± 30 g. Two 1m3 tanks were used for infected fish (60 fish in each tank), and one tank (60 fish) was used as negative control . The fish were followed for 12 weeks. All tanks were monitored by the colour digital camera in underwater housing during the cultivation. The video clips, which lasted 30 minutes, were recorded for all three tanks. These video records were used to simulate the detection of disease symptoms under real conditions. The dataset of the still images was created from five fish photographing sessions (every two weeks starting from week six). An image of the right and left side of the fish was captured for all fish. The expert scored the levels for oedema, scales, color, and RMS severity for each fish .
For the automatic detection of the RMS symptoms, six classes were defined to cover different skin deviations that belong to the RMS and other diseases. Three classes (early symptoms, late symptoms, and healing) were defined for RMS. Three classes were defined for RMS on fins, snout/lower jaw and other symptoms. All images were manually annotated to prepare the data for CNN-based RMS symptom detection. The CNN with YOLOv8 architecture was used for the detection and classification of annotated symptoms without the detection of individual fish .
The accuracy (mAP50) of RMS symptoms detection (still images) using the CNN approach was 87% (84% for early symptoms and 89% for late symptoms) . The early symptoms were detected by the system on the same day as the expert detected them . The study proved that the early symptoms of RMS disease can be automatically detected by deep learning methods from the still images of the fish. The next step is the RMS detection in video records to demonstrate the ability to detect the symptoms earlier than the expert.