A stroke image recognition model based on 3D residual network and attention mechanism

Abstract

In recent years, the number of stroke patients in China has been increasing and the development trend is not optimistic. In order to reduce the burden of doctors, improve the efficiency of clinical diagnosis and reduce the medical cost, the development of cerebral apoplexy imaging diagnosis is an inevitable trend. Taking stroke lesions in medical images as the object, a deep learning model 3D-SE ResNet10 is proposed which can distinguish whether stroke lesions are included in a given medical image with high accuracy. This model combines the attention mechanism with the residual learning network, and uses 3D convolution kernel to utilize the continuous information between slices in the medical image sequence. The model achieves an average accuracy of 88.69%, an average sensitivity of 87.58% and an average specificity of 90.26% in multiple experiments based on the realistic dataset. Its classification effect is significantly higher than that of 2D convolutional neural networks and 3D convolutional neural networks without attention mechanism. The experimental results show that our model is effective and feasible, and has certain practical value.