RFCBAM UNet: Enhancing UNet with Receptive-Field Convolutional Block Attention for Brain Tumor Segmentation Based on MRI

Brain tumor is a common type of tumor, which is extremely hazardous to human health. MRI is the most widely used non-invasive imaging tool for diagnosing brain tumors. Accurate segmentation of lesion regions from MRI is of great significance for brain tumor diagnosis and treatment planning. However, existing automatic segmentation algorithms, such as UNet and Swin-Unet, rely on a large amount of manually annotated data and have poor generalization on independent test datasets. While interactive segmentation algorithm is an important way to improve existing automatic segmentation algorithms by modeling the dynamic change process of image annotation and significantly improving the accuracy of model segmentation with better robustness and generalization. Therefore, this study focuses on the design of interactive brain tumor segmentation models based on multipoint prompt. First, in order to effectively integrate the multipoint prompt coding information with the multimodal MRI images information, this study proposes the ConvM module, which is able to better handle 4-channel multimodal MRI images. Second, in the part of feature extraction backbone network, this study proposes a new backbone network RFCBAM Unet which stands for Receptive-Field Convolutional Block Attention Module based on UNet. RFCBAM UNet improves the ability of the network to perceive channel information and spatial information, as well as solves the problem of parameter sharing in the process of attentional computation. Finally, this study supplementally employs the Object Contextual Representations (OCR) module to enhance the contextual representation of the tumor ROI pixels, thus improving the segmentation effect. After a series of experiments, it is shown that the RFCBAM UNet designed in this study is able to segment the tumor region of brain tumor patients more accurately with less number of interactions. The model can achieve a Dice coefficient of 0.8, 0.85 and 0.90 in an average of 3.43, 4.61 and 6.74 clicks, which is a series of facts that validate the usability and effectiveness of the model proposed in this study.