Résumé:
Skin diseases in animals are a growing concern in veterinary medicine, particularly in rural
and underserved areas that lack access to specialized care services. With the rapid
advancements in computer vision and artificial intelligence techniques, novel methods have
emerged to facilitate the early and accurate diagnosis of these pathological conditions.
This research proposes a predictive approach based on deep learning techniques, utilizing
Convolutional Neural Networks (CNNs) and multispectral imaging to detect and classify skin
diseases in dogs. The work is divided into three main axes: the first axis covers the
fundamentals of animal skin diseases, their clinical importance, and traditional diagnostic
methods. The second axis discusses the role of Transfer Learning and the use of state-of-the-
art neural network architectures in medical image classification. The third axis details the
experimental protocol, which includes the use of multispectral images, the class_weight
technique to address data imbalance, and the evaluation of several CNN models such as
Xception, EfficientNet, and DenseNet121.
The results demonstrated that the Xception model outperformed the other models, achieving
an accuracy of 98.9%, an Area Under the Curve (AUC) of 99.7%, and a Matthews
Correlation Coefficient (MCC) of 98.4%. These findings highlight the efficacy of combining
multispectral imaging with deep learning techniques in veterinary diagnostics and underscore
the significant potential of AI-powere d tools in enhancing animal health, especially in
resource-limited regions.
