AI-Driven Medical Image Segmentation for Early Tumor Identification in MRI Scans

Abstract

In this paper, a Hybrid Residual Attention Transformer U-Net (HRAT-UNet) model is proposed to solve the problem of accurate and interpretable brain tumor segmentation. In the proposed framework author have implemented Adaptive Frequency-Spatial Normalization (AFSN) with the aim of improving tumor-specific spectral features and Topology-Aware Morphological Filtering (TAMF) with the aim of maintaining the irregular shape of tumors in the pre-processing stage. Moreover, the residual convolutional feature extraction is embedded into the transformer based contextual encoding, thus better preserving the information of local structure as well as long-range spatial dependencies. Attention-guided skip connections are used for the delineation and segmentation precision. Spatial explanation maps are created using Grad-CAM and SHAP-based voxel attribution methods to make the model translational and better understood for clinical use. The proposed framework was tested with a publicly available Brain Tumor MRI dataset with a stratified train–validation–test protocol and data augmentation techniques. The experimental results showed that the proposed model outperforms the traditional CNN, U-Net, Attention U-Net and Vision Transformer based segmentation models with 98.76% accuracy, 98.21% Dice score and 97.05% Intersection over Union (IoU). The results show that the developed HRAT-UNet architecture is an effective and interpretable method to early detection and segmentation of tumors in the brain from MRI sequences during clinical applications.