A Medical Image Intelligent Recognition Method

Abstract

This invention relates to the field of medical image classification, specifically to an intelligent medical image recognition method. This invention employs transfer learning to improve the feature extraction performance of the backnone network and accelerates training. Non-local modules are used to capture long-distance dependencies. Within these modules, parallel and cascaded convolutions of different sizes and dilation rates are used to obtain the key and query, capturing fine-grained texture. Finally, global average pooling is used to fuse spatial features, and 1x1 convolutions are used for prediction, with Softmax used to activate the prediction results. This invention accelerates training, captures fine-grained texture, is more conducive to medical image classification, improves classification accuracy, and provides assistance to clinical practice.

Images