U-Net++DSM: Improved U-Net++ for Brain Tumor Segmentation With Deep Supervision Mechanism

The segmentation of brain tumors is an important and challenging content in medical image processing. Relying solely on human experts to manually segment large volumes of data can be time-consuming and delay diagnosis. To address this challenge, researchers have set out to develop an algorithm that can automatically determine whether MRI images contain brain tumors and identify their features. This paper proposes the U-Net++DSM, a collaborative approach combining U-Net++ with Deep Supervision Mechanism (DSM) as its backbone. To enhance the segmentation power of U-Net++DSM, medical professionals have trained a dilation operator using fully annotated images. The results of this method demonstrate that the combination of U-Net++DSM and the dilation operator significantly improves segmentation accuracy, especially when the number of fully-labeled images is limited. The results show that the proposed U-Net++DSM outperforms traditional U-Net models by achieving high segmentation performance, surpassing other state-of-the-art models, with a sensitivity of 98.59%, a specificity of 98.64%, an accuracy of 98.64%, and an average Dice score of 98.02% when tested on publicly available databases. Compared to other existing segmentation methods, the U-Net++DSM method has the potential to yield even better brain tumor segmentation results in terms of pixel-based classification and dice similarity performance metrics.