AENet: End-to-end training of Attention Efficientnet for stroke segmentation

Medical image segmentation has extensive application and research value in medical research and practice fields, such as clinical diagnosis, pathological analysis, Operation Planning, image information processing, computer-assisted surgery, and so on. In the medical imaging of stroke, the lesions can be distributed in any area, and these lesions vary in size and shape. At present, various network frameworks have been used in deep learning for medical image segmentation, and good results have been achieved. However, the extremely small lesions in stroke images have serious neglect and false detection problems. To solve this problem, we design a loss function. This loss function optimizes the learning of semantic feature information around the target lesions, which can assist the network to locate small target regions more easily in the early stage of training. Meanwhile, it can add spatial information to assist the guidance network to achieve better convergence, accelerate network training, and finally improve the segmentation accuracy. In addition, we propose an end-to-end network structure, where the decoder adds the Attention Gates module to focus Attention on salient features that are useful for a particular task, the final prediction result is obtained by using the decoder structure to recover the spatial resolution. We sent the stroke images of different size lesions to the network for training by combining this network structure and the loss function we designed. The experimental results show that the DICE and NSD of our experimental data are improved effectively.