Optimizing multimodal MRI brain tumor segmentation via lightweight encoders and model pruning

Accurate segmentation of brain tumors from multimodal MRI data plays a pivotal role in neurooncological diagnosis, therapeutic strategy development, and prognostic assessment. While contemporary deep learning approaches deliver exceptional accuracy, their computational intensity often limits practical implementation in clinical environments, particularly on devices with restricted resources. This study confronts this limitation by developing an efficient segmentation framework that balances computational economy with diagnostic precision through streamlined encoder designs and systematic network pruning. Utilizing the BraTS 2024 post-treatment glioma benchmark a specialized dataset capturing treatment-induced anatomical changes this work investigates the replacement of traditional convolutional encoders with MobileNet-derived architectures employing depthwise separable convolutions. These architectural modifications substantially decrease parameter quantities and computational operations while preserving segmentation fidelity. A complementary pruning strategy selectively removes non-essential network components to enhance operational speed and reduce memory footprints. The optimized segmentation system maintains competitive performance across critical clinical metrics (Dice similarity coefficients) while achieving notable advancements in processing speed iii and resource efficiency. When evaluated against the most recent models, the proposed framework demonstrates equivalent segmentation accuracy with approximately 40inference capabilities.