Deep-Feature Fusion and Hybrid Machine Learning for Canine Dermatological Image Classification

Automated classification of canine dermatological images is challenging because visually overlapping lesions, uncertain repository labels, and repeated or transformed source photographs can influence performance estimates. This study evaluated fixed ImageNet-pretrained DenseNet201, ResNet50, and EfficientNetB0 features, alone and in fusion, with support vector machine (SVM), multilayer perceptron (MLP), and logistic regression classifiers for six repository-provided categories. Transformation-robust source auditing yielded 3088 images: 2862 development images from 864 source groups and a source-group-separated held-out internal test set of 226 images from 226 groups. Model selection used five-fold source-group-aware outer cross-validation with source-group-aware inner grid search. DenseNet201–EfficientNetB0 + SVM achieved the highest mean outer-fold balanced accuracy (0.9462 ± 0.0174), while triple fusion was comparable (0.9449 ± 0.0142) with a larger feature representation. On the held-out internal test set, accuracy was 0.8319 (95% CI 0.7832–0.8805), balanced accuracy was 0.7806 (0.7219–0.8384), and macro F1 was 0.7871 (0.7234–0.8432). Independent, clinically annotated external validation remains necessary to assess generalizability. These findings highlight the value of source-aware evaluation and parsimonious feature fusion when assessing multiclass canine dermatological image classification using repository-based datasets.