A Dual-Attention and Uncertainty-Aware Deep Learning Framework for Clinically Trusted Brain Tumor Segmentation

Abstract

Although brain tumour segmentation is essential for medical image analysis, reliable and accurate segmentation with clinical and routine quality appears to be difficult because of the heterogeneity of the tumour, the lack of clear boundaries and the noise within the images. Current deep learning (DL) models have been plagued by the problem of the black box effect and overconfident predictions, limiting their use in real clinical settings. In fact, current DL models have experienced the issue of black box effect and high prediction accuracy, which prevent the models from being used in the real clinical setting. To solve these, the study puts forward a Dual-Attention and Uncertainty-Aware DL framework to enhance segmentation accuracy and clinical trustworthiness. The method combines both spatial and channel attention mechanism and is implemented in an encoder-decoder framework which allows obtaining more informative features from multi-modal MRI scans. Besides, a multi-scale feature fusion module is used to capture tumors of different scales, and an uncertainty-aware module is used to measure the confidence of prediction by computing the stochastic inference. A clinical trust calibration layer is added to make the predicted probabilities more realistic. Experiments performed on the BraTS dataset show better performance when compared with the state-of-the-art models with Dice score, IoU, sensitivity, and specificity of 94.8%, 89.5%, 93.6%, and 97.1%, respectively. The proposed framework not only enhances the segmentation accuracy but also generates the uncertainty maps, which are well suited for a clinical decision support system. Finally, the model provides a reliable, interpretable, and strong brain tumor segmentation solution, which has great potential in medical applications.