MedX-Net: Hierarchical Transformer with Large Kernel Convolutions for 3D Medical Image Segmentation

Due to the exceptional performance of Transform-ers in 2<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$D$</tex> medical image segmentation, recent work has also introduced them into 3D medical segmentation tasks. For instance, Swin UNETR and other hierarchical Transformers have reintroduced prior knowledge from several convolutional networks, further enhancing the volume segmentation capabilities of models. The efficacy of these hybrid methodologies is primarily attributed to the substantial quantity of parameters and the nonlocal self-attention mechanism with a large receptive field. We argue that the behavior of these methods' large receptive fields can be simulated by employing fewer parameters through the utilization of depth-wise convolutions with large kernel. Within this manuscript, we introduce a lightweight volume segmentation model called MedX-Net, which uses convolutional network modules to simulate hierarchical Transformers for robust volume segmentation. Firstly, inspired by the hierarchical Transformer module of Swin UNETR, we investigate large-kernel depth-wise convolutions with different sizes to achieve a reduced model parameter count while maintaining a large global receptive field. Secondly, we replace the multilayer perceptron (MLP) in the hierarchical Transformer module with Inverted Bottleneck with Depthwise Convolution Enhancement(DWCE) to improve model performance with fewer activation and normalization layers, further reducing the parameter count. We validate the effectiveness and efficiency of our model for volume segmentation on three public datasets: Synapse, BTCV and ACDC. On the Synapse dataset, compared to Swin UNETR, our model achieves an improvement from 83.48% to 87.21% in Dice score. Compared to the result of 86.57% achieved by nnFormer, our model achieves superior performance while reducing the model parameter count by 64%.