BTIS-Net: Efficient 3D U-Net for Brain Tumor Image Segmentation

Brain tumor segmentation techniques are essential for the precise delineation of tumors and normal brain tissues which is essential for the guidance of surgical intervention and clinical decisions. However, for resource-constrained clinical environments, more efficient and lightweight segmentation models are needed so that they can be applied in real-time for surgical navigation and clinical decision-making. To tackle this issue, the proposed study introduces a very effective 3D U-Net model that is specifically designed for brain tumor image segmentation. This study presents the primary contributions as follows: 3D depth separable convolution is introduced to decrease the number of model training parameters, hence enhancing the overall efficiency of the model. The dilated dense residual block is designed to expand the sensory field, allowing the network to grasp a broader range of features and structures within the input data. As a result, the model’s performance and generalization ability to handle complex tasks are improved. The integration of the confusion area segmentation module enhances the model’s capability to discern intricate image details and edges, thereby augmenting the overall segmentation efficacy. Evaluation of the proposed BTIS-Net involves experimentation on two widely recognized datasets, namely BraTS 2019 and BraTS 2021. The Dice similarity coefficient, Positive predictive value, and Sensitivity exhibit average improvements of 5.68, 5.38, and 2.14, respectively. Additionally, the Hausdorff distance is reduced by an average of 2.71. The experimental results validate the efficient segmentation performance of the BTIS-Net model, showcasing exceptional outcomes even under resource constraints.