Efficient brain tumor segmentation using Hybrid-Pruned nnU-Net

Accurate medical image segmentation plays a crucial role in diagnosing diseases and planning effective treatments, particularly in detecting brain tumors using Magnetic Resonance Imaging (MRI) scans. These segmentations provide detailed insights required to identify problems and deliver timely and effective care. The contemporary models, like nnU-Net, provide exceptional accuracy. However, their processing requirements make it challenging to reach real-time performance. This study proposes an efficient feature map pruning method that concurrently increases nnU-Net’s segmentation accuracy and reduces inference time. Several pruning criteria are explored to enhance the model’s efficiency, such as the variance, mean activation, and a novel hybrid approach that combines activation statistics with weight norms. Through the assessment of these various approaches, we determine the best strategy for reducing unnecessary complexity without compromising model functionality. This process is further refined by introducing a voting mechanism that acts like a panel of experts, selecting the best pruning decisions based on multiple criteria. This collaborative approach ensures more reliable and consistent outcomes, optimizing the model’s accuracy and latency while maintaining its streamlined and efficient nature. On the BraTS 2023 dataset, the pruned nnU-Net models achieve inference times of 0.0406 s, compared to 2.9258 s for the original model, which is a considerable improvement. Our pruned models maintain competitive Dice scores of up to 0.7585. These results demonstrate the potential of our approach for time-sensitive medical image processing.