Abstract
<jats:p>A new DSP-enhanced deep learning approach for brain tumor MRI classification and segmentation is proposed in this paper. Our approach includes a Learnable Wavelet CNN (LW-CNN) for adaptive frequency-aware denoising, a differentiable anisotropic diffusion layer for edge-preserving diffusion, and a Hierarchical Multi-Scale Deformable Attention Module (MS-DAM) for feature extraction. A U-Net with attention mechanism is used to perform accurate segmentation in an integrated multi-task learning framework. The proposed framework is different from traditional methods that only use deep feature learning. It uses digital signal processing methods to improve the quality of the input before learning how to represent it. The system is also made to work at the edge, with preprocessing done on FPGA and inference done on NVIDIA Jetson platforms. Experimental data on a multi-class brain tumor MRI collection show that, in terms of accuracy and Dice score, the suggested technique outperforms state-of-the-art CNN, attention, and transformer-based models. Noise tolerance analysis further validates the performance of the suggested DSP components in enhancing picture quality and structural preservation.</jats:p>