Depthwise Separable and Multi-Scale Dilated Convolution Residual-Unet for Left Atrial Segmentation

다국어 초록 (Multilingual Abstract)

Convolutional Neural Networks (CNNs) have been proven to be powerful models for medical image segmentation, and this study aims to address some of the major unsolved problems in these models, such as a small image dataset, sample im-balance, and a small target sample. This paper proposes an improved depthwise separable and multi-scale dilated ResUnet (DSMSD-ResUnet) based on Residual-Unet by combining the ideas of residual connection, depthwise separable convolution (DSC) and multi-scale dilated convolution (MSDC): (1) Based on the structure of the Residual-Unet, the DSC is used to replace the standard convolution, and the model parameters are reduced without changing the segmentation effect. (2) The deep-level features extracted by the depthwise separable residuals are added to the multi-scale dilated module to enhance the edge discrimination ability of the image block. (3) Introducing an integrated attention mechanism (AM) into the network to suppress the interference of non-left atrial regions with the segmentation results of the left atrial region. (4) An improved loss function is introduced for imbalance and small target data samples. The experimental results demonstrate that the proposed method can effectively enhance segmentation performance, and the average Dice close to 0.9 is obtained on public dataset, and the results of other evaluation indicators are also satisfactory. This paper illustrates the accuracy of the learning method in the current segmentation task.