FSTU-Net: volumetric fast convolution and hierarchical transformer for faster medical image segmentation

Recently, hybrid CNN-Transformer models have attained a state-of-the-art performance in 3D medical segmentation tasks. The effectiveness of this approach is predominantly ascribed to the large receptive field for non-local self-attention, albeit at the cost of a substantial number of model parameters. Despite the introduction of deep convolution to reduce FLOPs of the model, deep convolution still has the problem of frequent memory access. In this paper, we introduce FSTU-Net, a lightweight 3D medical image segmentation model, which aims to adapt the inherent features of the Swin Transformer through ConvNet modules and enhance volumetric segmentation performance using a smaller model capacity. Specifically, we introduce Volumetric Fast Convolution (FConv) with Large Kernel (LK) size to emulate the large receptive field operations generated from attention in the Swin Transformer. We have developed a simple, rapid, and efficient new Fast Convolution (FConv) to eliminate the problem of frequent memory access associated with deep convolution or group convolution, reducing computational redundancy and the number of memory accesses. It holds significant potential to replace the current preferred method of deep convolution. We validated our FConv method and FSTU-Net model on BTCV and AMOS 2022 public datasets, assessing both efficiency and accuracy. Our results demonstrated state-of-the-art performance, effectively reducing the number of model parameters.