Self-supervised wavelet-based attention network for semantic segmentation of MRI brain tumor

Abstract

To determine the appropriate treatment plan for patients, radiologists must reliably detect brain tumors. Despite the fact that manual segmentation involves a great deal of knowledge and ability, it may sometimes be inaccurate. By evaluating the size, location, structure, and grade of the tumor, automatic tumor segmentation in MRI images aids in a more thorough analysis of pathological conditions. Due to the intensity differences in MRI images, gliomas may spread out, have low contrast, and are therefore difficult to detect. As a result, segmenting brain tumors is a challenging process. In the past, several methods for segmenting brain tumors in MRI scans were created. However, because of their susceptibility to noise and distortions, the usefulness of these approaches is limited. Self-Supervised Wavele- based Attention Network (SSW-AN), a new attention module with adjustable self-supervised activation functions and dynamic weights, is what we suggest as a way to collect global context information. In particular, this network’s input and labels are made up of four parameters produced by the two-dimensional (2D) Wavelet transform, which makes the training process simpler by neatly segmenting the data into low-frequency and high-frequency channels. To be more precise, we make use of the channel attention and spatial attention modules of the self-supervised attention block (SSAB). As a result, this method may more easily zero in on crucial underlying channels and spatial patterns. The suggested SSW-AN has been shown to outperform the current state-of-the-art algorithms in medical image segmentation tasks, with more accuracy, more promising dependability, and less unnecessary redundancy.