LDS ‐ UNet : Lightweight Depthwise Separable U‐Net for Brain Tumor Segmentation

Brain tumor segmentation from MRI plays an important role in clinical decision making. Conventional U‐Nets incur high computational cost that limits real‐time use. To mitigate this issue, we propose a hybrid Lightweight Depthwise Separable U‐Net (LDS‐UNet) that uses depthwise separable convolutions in the encoder‐decoder and standard convolutions in the bottleneck, and a three‐slice stacked (2.5D) input to recover inter‐slice context without the expense of full 3D convolutions. Independent binary models are trained for Whole Tumor (WT), Tumor Core (TC), and Enhancing Tumor (ET) using a compound Dice‐Focal‐IoU loss. We evaluate the method on BraTS 2020 (369 patients, four modalities) using a patient‐wise split (209 training, 37 validation, 123 test volumes), with metrics computed at the patient level. Across three seeds(42,43,44), mean Dice reaches 0.8762 (WT), 0.7860 (TC), and 0.7139 (ET), with ROC‐AUC of 0.976, 0.938, and 0.951. Compared to a Traditional U‐Net baseline with the same protocol, the model achieves 40% less parameters (4.66 M vs. 7.77 M) and 67% less computation (4.0 vs. 12.2 GFLOPs). Controlled ablation confirms the hybrid bottleneck, compound loss, and 2.5D input each contribute positively, with 2.5D input giving significant improvements on TC/ET over single slice input. The entire pipeline takes an estimated 3.78 s per patient volume on a Tesla T4 GPU. Building on the current single patient‐wise split, future work will incorporate a controlled 3D/transformer baseline, along with lightweight attention, multi‐center validation, and edge quantization to further strengthen clinical applicability.