Disease in fish is a major threat to aquaculture, threatening fish health, productivity, and economic sustainability. Traditional diagnosis is subjective and time-consuming visual inspection, which underscores the need for better alternatives. This work proposes and compares advanced AI-based fish disease categorization models based on transfer learning with four of the top-performing pre-trained CNNs: ResNet50, DenseNet121, InceptionV3, and MobileNetV2. These models are trained on seven-class fish disease with strong data augmentation and regularization to enhance generalization. A new Performance Metric-Infused Weighted Ensemble (PMIWE) approach is also enhanced, with metric-based models that are weighted adaptively based on metrics like precision, recall, F1-score, MCC, and Cohen’s Kappa. The PMIWE ensemble outperforms base models and the averaged ensemble, with test accuracy of 92.86% and high AUC. Large-scale examination with classification reports, confusion matrices, and ROC analysis establishes the power of the method and the feasibility for future research into model interpretability and real-time aquaculture monitoring.