MLKCA-Unet: Multiscale large-kernel convolution and attention in Unet for spine MRI segmentation

Abstract

Medical image segmentation plays a key role in the diagnosis of spinal diseases. Unet has become a universal structure for image segmentation because of its unique skip connection structure in recent years. However, since Unet uses small-kernel convolution, the relationship between remote features is difficult to obtain due to the small receptive fields, and the key information cannot be highlighted, resulting in insufficient edge information. To overcome these problems, this paper proposes multiscale large-kernel convolution Unet (MLKCA-Unet), which develops MLKC block for effective feature extraction. Large-kernel convolution with different convolution kernels is used according to the feature map. For large feature maps, smaller large- kernel convolution is used, and for small feature maps, larger large-kernel convolution is used. All large-kernel convolution can be reduced the dimension by 1 × 1 convolution kernel. This method has a significant reduction in computation. By paralleling each large kernel convolution branch with the 3 × 3 convolution branch, it helps to capture detailed information. At the same time, an attention mechanism is added to the network to emphasize rich feature areas and enhance useful information. Finally, various indicators are employed to evaluate the network’s accuracy, similarity and speed, including IOU, DSC, TPR, PPV, and ET. The published spinesagt2wdataset3 spinal MRI image dataset is adopted in the experiment. The IOU, DSC, TPR, PPV, and ET on the test set are 0.8302, 0.9017, 0.9000, 0.9051 and 70 s/epoch respectively. The experimental result shows that MLKCA-Unet demonstrates superior segmentation performance and robustness, which can be well extended to other medical image segmentation.