Lesion-Preserving and Confidence-Aware Fish Lesion Segmentation for Sustainable Aquaculture and Aquaponic Health Monitoring

Timely fish disease monitoring is an important requirement for sustainable aquaculture because disease outbreaks can reduce survival, increase treatment inputs, and destabilise production. In aquaponic systems, fish health is also linked to nutrient cycling and the stability of integrated fish–vegetable production, making automated fish-health perception a potentially useful component of resource-efficient farming. Existing classification and detection methods can identify disease categories or approximate lesion locations, but they provide limited information about lesion area, boundary shape, and severity-related spatial extent. This study presents a deep learning framework for pixel-level fish lesion segmentation to support sustainable aquaculture health monitoring, with aquaponic systems considered as a potential application context. The framework combines lesion-preserving frequency augmentation (LPFA), confidence-guided large-kernel encoding (CGLE), and confidence-filtered decoder refinement (CFDR). LPFA expands lesion appearance variation during training while retaining the main lesion layout. CGLE uses coarse prediction confidence to allocate broader contextual modelling to uncertain encoder regions, and CFDR applies selective decoder correction to low-confidence regions. A public freshwater fish disease dataset is reformulated into a dense prediction task with 1750 raw images from seven image-level categories, including six disease categories and one normal healthy category. The images are divided into training, validation, and test subsets at an 8:1:1 ratio, and controlled augmentation strategies are applied online rather than being used to create a larger static dataset. Across five random-seed runs, the proposed method achieves 82.6±0.3% mIoU, 90.9±0.2% mDice, and 73.5±0.4% Boundary IoU. Relative to TransUNet, the mean mIoU rises from 78.4±0.4% to 82.6±0.3%, and Boundary IoU rises from 68.8±0.5% to 73.5±0.4%, with paired bootstrap testing supporting the stability of the improvement. These results indicate its potential as a lesion-quantification decision-support component for smart and sustainable fish-production systems.