WAVE-VNET: Enhancing Medical Image Segmentation with Wavelet Transformer

Transformer-based models, renowned for their ability to learn long-range information, have recently shown outstanding ca-pabilities in the fields of computer vision and medical image analysis. However, a common limitation in these models is their tendency towards excessive down-sampling, which often results in the loss of high-frequency information. To over-come this limitation, this paper proposes a new method, called Wave-VNet, for efficient and accurate segmentation for medi-cal images. This method employs tokenized image patches in a Transformer-based, V-shaped Encoder-Decoder structure, enriched with skip connections to enhance local-global se-mantic feature learning. Taking advantage of Wavelet Trans-form for invertible down-sampling using the Discrete Wavelet Transform (DWT) which preserves the original image details while minimizing computation, it integrates Wavelet Trans-former as the encoder, structured into three stages, which is designed for extracting contextual features, while a sym-metrically designed Wavelet Transformer-based decoder, featuring patch expanding layers, is employed for efficient up-sampling, thereby preserving the spatial resolution of the feature maps. We have examined Wave-VNet on Synapse and ACDC datasets. Our extensive experiments demonstrate that Wave-VNet outperforms various existing methods. It can achieve a 10% increase in the Dice coefficients and an over 18% increase in Hausdorff distance over the prior arts.