DWT-MFN: A Dual Swin Transformer-based Multi-scale Fusion Network with Autoencoder Regression for Brain Tumor Segmentation and Survival Prediction

A three-step brain tumor analysis method is introduced that includes pre-processing, segmentation, and survival prediction.The Unified Denoising and Enhancement Filter (UDEF) on a Statistical Wavelet Framework (SWF) is used to reduce noise in input MRI scan images of brain tumor tissue in the publicly available BRATS 2020 dataset without compromising structural information like the tumor tissue boundary.A DWT-MFN-based Multi-Scale Fusion Network divides tumors next.It uses a two-stream encoder with window-based and shifted-window selfattention to capture tumors of various sizes, textures, and shapes.Finally, a Stacked Sparse Autoencoder with Locally Weighted Regression (SSAE-LWR) can learn hierarchical representations over segmented tumor regions to automatically learn features to predict patient survival (tumor size, irregularity of boundaries, distribution).The proposed framework outperforms state-of-the-art models on the BRATS 2020 dataset with 99.42 accuracy, 99.35 sensitivity, and 99.37 specificity.AUC (0.9935), Hausdorff distance, Dice score (0.9051), and F1-score (0.985) demonstrate its strength and accuracy.