Application of Convolutional Neural Networks in the Detection and Classification of Fish Diseases using Image Analysis

Background: Early and accurate detection of fish diseases is essential for improving aquaculture productivity and reducing economic losses. Convolutional neural networks (CNNs) offer a promising solution for automated disease classification using image-based analysis, yet the impact of different data split strategies on model performance requires further investigation. Methods: A CNN-based approach was developed to classify common fish diseases using a dataset comprising Aeromoniasis, Gill disease, Healthy Fish, Parasitic diseases, Red disease, Saprolegniasis and White spot disease. Two data split ratios, 80:10:10 and 70:15:15, were evaluated. Model performance was assessed using accuracy, macro-average F1-score, confusion matrices, validation loss and ROC curves. Result: The 70:15:15 split achieved higher overall accuracy (86.46%) compared to the 80:10:10 split (82.81%) and demonstrated a superior macro-average F1-score (0.8474 vs. 0.8266). However, this configuration exhibited signs of overfitting, reflected by increased validation loss. In contrast, the 80:10:10 split showed more stable validation loss and reduced misclassification in certain disease categories. High precision was consistently observed across most classes, although recall varied due to class imbalance and overlapping visual features. The findings indicate that increasing the validation set size can enhance classification performance but may also introduce overfitting risks. Implementing appropriate regularization and data balancing techniques is therefore essential. The proposed CNN-based framework demonstrates strong potential for integration into automated aquaculture monitoring systems, enabling timely disease detection, minimizing losses and improving overall fish health management.