Deep Learning et l’Aide au Diagnostic Médical (Application à la Classification d’image médicale)

Abstract

Medical image classification has seen significant progress in recent years, driven by rapid ad- vancements in medical imaging techniques and the emergence of deep learning methods, which

have proven highly effective in the field of computer vision. Deep learning is characterized by its ability to automatically extract features from data without requiring complex preprocessing steps, enabling the construction of rich hierarchical representations of visual information. These developments have enhanced the efficiency of classification systems, making them better suited to handle the visual complexity inherent in many medical conditions, particularly in medical imaging diagnostics, such as diabetic retinopathy classification. This is precisely the focus of

our research. Accordingly, we designed an automated system to effectively classify diabetic re- tinopathy using deep learning networks, specifically leveraging transfer learning to ensure high

performance and reasonable computational time. We proposed three architectures based on the

convolutional models DenseNet121, ViT-B/16, and YOLOv8, which are renowned for their ef- fectiveness on the ImageNet dataset. Our contribution involves classification using two distinct

strategies : Direct classification, where the model is applied directly to the images. Cascade classification, where images are classified in stages—first into 2 classes (binary), then 3 classes, and finally 5 classes. The experiments were conducted on two datasets : The first dataset used

is APTOS2019, comprising 3,662 images distributed across five (5) distinct classes. The results demonstrated a clear superiority, achieving a maximum accuracy of 96.30%. The second dataset is the DDR (Dataset for Diabetic Retinopathy), consisting of 12,522 images also divided into five (5) distinct classes. The results were satisfactory, with dataset accuracy reaching 93% and 96% respectively.