Deep Learning pour l’analyse et la classification d’images

Abstract

In the medical eld, image classi cation has achieved great success following advances in imag ing technologies and deep learning, which is considered the most e ective tool in visual recognition by automating feature extraction and developing hierarchical representations of structures and pat terns, thereby enhancing the e ciency of classi cation methods. In this context, we have developed an automatic system aimed at e ciently classifying peripheral blood cells by leveraging Deep Learn ing networks through the use of transfer learning techniques for better performance and reasonable computation time. Three designs were proposed based on the convolutional architectures VGG16, InceptionV3, and ResNet50, known for their high performance on the ImageNet dataset. The rst implementation is a combination of VGG16 and InceptionV3, the second is a concatenation of Incep tionV3 and ResNet50, and the third is an association of VGG16 and ResNet50. Experiments were conducted on the "Peripheral Blood Cell" (PBC) dataset, consisting of 17 092 images divided into eight distinct classes. The results obtained were very satisfactory, with a maximum accuracy of 99%.