MRFB-Net: A Novel Attention Pooling Network With Modified Receptive Field Block for Uterine Fibroid Segmentation

The segmentation of uterine fibroid in ultrasound image is very important for the implementation of ultrasound-guided high-intensity focused ultrasound therapy. However, due to the inherent limitations of ultrasound imaging, such as poor contrast, speckled noise, and variability of fibroid appearance, reliable and effective segmentation algorithms still face great challenges. To overcome these limitations, we introduce a deep learning-based architecture that leverages attention pooling decoder module to enhance the segmentation of uterine fibroids in preoperative ultrasound images, named MRFB-Net. In our framework, we utilize an encoder-decoder structure inspired by U-Net to capture image features. Then, we design an attention pooling module (APM) into the decoder to effectively alleviate the loss of spatial details caused by multiple pooling operations. Subsequently, we integrate modified receptive field block (MRFB) into the skip connection to learn more robust feature representation. Furthermore, we introduce a multilayer feature fusion module (MFFM) to harness complementary information from different layers. Finally, we conducted a series of comparative experiments using two in-house uterine fibroid clinical datasets and one publicly available brain tumor dataset. The evaluation results demonstrate the accuracy and reliability of the proposed MRFB-Net, as well as its adaptability to different types of medical images. In addition, ablation experiments also verified the effectiveness of the attention pooling module, modified receptive field block and multilayer feature fusion module.