TY  - EJOU
AU  - Patel, Ronak 
AU  - Patel, Miral 
AU  - Kothadiya, Deep 
AU  - AlGhofaily, Bayan 
AU  - Alamri, Faten S. 
AU  - Alyousef, Awad 
AU  - Khan, Amjad R 

TI  - LiteDKT-Net: A Lightweight Diverse Kernel Transformer Network for Brain Tumor Segmentation
T2  - Computers, Materials \& Continua

PY  - 
VL  - 
IS  - 
SN  - 1546-2226

AB  - Growth of cancerous cells is unpredictable, and their effects vary across organs and levels of aggression. Identification of the pattern, size, and shape of the growth helps assess severity for better treatment. The proposed LiteDKT-Net combines the DK-IRB (Diverse Kernel Inverted Residual Block) block and Transformer to target conceptual information about shape and location. For better edge detection, LiteDKT-Net uses GAG (Group Attention Gate) followed by CBAM (Convolutional Block Attention Module). LiteDKT-Net is a lightweight encoder-decoder-based network optimized for accurate brain tumor segmentation. The network parameter optimization and reduced computational complexity in LiteDKT-Net enable high segmentation accuracy while maintaining a lightweight model size for deployment in clinical environments with limited resources. The proposed architecture achieves Dice score similarities of 0.80, 0.82, and 0.87 for ET (Enhancing Tumor), TC (Tumor Core), and Whole Tumor (WT), respectively. To check the difference between the predicted and ground truth with HD95, the results are 5.24, 8.12, and 8.01 mm for ET, TC, and WT. The proposed architecture evaluates 0.869M parameters and uses 0.856 Giga Floating-Point Operations Per Second (GFLOPs) of computation. Ablation analysis is carried out based on the channel-wise and kernel-wise parameters, which helps to understand model efficiency and lightweight computational cost.
KW  - UNET; transformer; group attention gate; CBAM; 3D MRI segmentation; FLOPs

DO  - 10.32604/cmc.2026.085703