Early Brain Tumor Segmentation using Hybrid Deep Learning and Kolmogorov–Arnold Networks

Abstract

Primary brain tumors are glioblastoma multiforme (GBM) and lower-grade gliomas (LGGs) which are very fatal. Clinically, multi-parametric MRI is still a challenge in accurate and early stage localisation of subregions of tumours. Existing automated algorithms have difficulties in sensitivity to boundaries, cross protocol generalization, and long-range global context.

We present CNN-Mamba-KAN which is a hybrid deep learning framework that combines three paradigms in order to have a robust brain tumor segmentation strategy. A ResNet-50 convolutional encoder projects hierarchical local textures of three spatial scales, and parallel Mamba-2 state-space model (SSM) branches are effective to project global volumetric dependencies with linear complexity. These representations are entailed through cross-attention fusion gates. This architecture has been designed with attention-enhanced multi-resolution bottleneck based on a variant of Convolutional Block Attention Module (CBAM). More importantly, the decoder uses B-spline activations that are learnable and implemented as Kolmogorov-Arnold Network (KAN) layers instead of multi-layers perceptron in an expression of features. Training was done on a compound Dice, focal, and boundary loss to overcome class imbalance and improve boundary delineation.

CNN-Mamba-KAN was evaluated with state-of-the-art performance on the dataset BraTS 2024 by five-fold cross-validation. The Dice Similarity Coefficients (DSC) and 95% Hausdorff Distances (HD95) in the model stood at 92.4% and 3.21mm, respectively, of Whole Tumor, Tumor Core, and Enhancing Tumor. These findings show statistically significant gains (p < 0.05) compared to modern models such as EfficientMed and MambaND and scale-effectively (48.3M parameters, 1.8 seconds/volume inference).

Through convolutional local extraction, SSM-based global context, and KAN-based adaptive decoding, CNN-Mamba-KAN is able to enhance the process of critical subregion segmentation of the tumors. Hence, it provides an extremely helpful, automated, clinical MRI treatment planning and patient diagnosis tool.