<![CDATA[Canine skin diseases represent a significant health concern in veterinary practice, with accurate diagnosis often requiring specialized expertise and microscopic examination. This study presents an implementation and evaluation of the ConvNeXt architecture for classifying microscopic images of four canine skin diseases: Demodex, Kokus, Malassezia, and Scabies. The dataset consists of microscopic images obtained from skin scraping preparations photographed under 400× digital microscopy and annotated by a certified veterinary. Two dataset scenarios were evaluated: a small dataset (356 images; Demodex: 104, Kokus: 47, Malassezia: 102, Scabies: 103) and a large dataset (2,963 images with near-balanced class distribution). Using transfer learning with pre-trained weights from ImageNet, the ConvNeXt model was evaluated across three input sizes (224×224, 180×180, 150×150). Augmentation balancing, including rotation, flipping, zoom, translation, shear, and color jitter, was applied to address class imbalance while preserving biological validity of morphological features Augmentation balancing was applied to address class imbalance, ensuring equal representation across all classes. Experimental results demonstrate that ConvNeXt with 224×224 input size trained on the large dataset achieved the best overall performance with 97.22% test accuracy, 0.9524 F1-score, 0.9624 Matthews Correlation Coefficient (MCC), and a perfect Area Under Curve (AUC) of 1.0000. Analysis of input size effects revealed that 224×224 is optimal for detecting small pathogens like Malassezia (3-8 μm) and Kokus (0.5-1 μm), while 150×150 better preserves spatial context for large pathogens such as Demodex (300 μm) and Scabies (200-400 μm). Visualization of feature maps provided insights into how the architecture extracts diagnostic features, producing dense, hierarchical feature representations that benefit from abundant data. This research demonstrates that transfer learning with the ConvNeXt architecture, combined with appropriate augmentation balancing, can achieve high classification accuracy for automated diagnosis support of canine skin diseases. However, clinical deployment requires further validation by veterinary and prospective clinical studies before these results can be considered clinically applicable.]]>
