Abstract

Info

Title:: Multi-scale attention medical image segmentation network fused with wavelet transform

Author(s):: ZHU Gang; LI Sihai; School of Medical Information Engineering, Gansu University of Chinese Medicine, Lanzhou 730000, China

Keywords:: Keywords: medical image segmentation encoder-decoder structure discrete wavelet transform multi-scale attention

Abstract:: Abstract: The semantic gap between the encoder and the decoder in image segmentation models, along with noise interference during feature propagation, restricts the segmentation accuracy. To address this issue, a WMDUNet model is constructed, which adopts discrete wavelet transform to downsample feature maps in the encoder, extracting low-frequency feature components and removing high-frequency feature components for noise elimination. Additionally, multi-scale dual attention is used to enhance the skip connection, bridge the semantic gap between encoders and decoder, and enable high-precision medical image segmentation. On the ACDC and PROMIS12 datasets, this method achieves the best results in 3 evaluation metrics, with mean Dice similarity coefficients of 91.46% and 85.85%, mean intersection over union of 87.64% and 75.22%, and mean Hausdorff distance of 1.22 and 3.51 mm, respectively, which verifies its effectiveness.