MAF-net: multi-receptive attention fusion network with dual-path squeeze-and-excitation enhancement module for uterine fibroid segmentation.

Introduction: Uterine fibroids are one of the most common benign tumors affecting the female reproductive system. In clinical practice, ultrasound imaging is widely used in the detection and monitoring of fibroids due to its accessibility and non-invasiveness. However, ultrasound images are often affected by inherent limitations, such as speckle noise, low contrast and image artifacts, which pose a substantial challenge to the precise segmentation of uterine fibroid lesions. To solve these problems, we propose a new multi-receptive attention fusion network with dual-path SE-enhancement module for uterine fibroid segmentation.

Methods: Specifically, our proposed network architecture is built upon a classic encoder-decoder framework. To enrich the contextual understanding within the encoder, we incorporate the multi-receptive attention fusion module (MAFM) at the third and fourth layers. In the decoding phase, we introduce the dual-scale attention enhancement module (DAEM), which operates on image representations at two different resolutions. Additionally, we enhance the traditional skip connection mechanism by embedding a dual-path squeeze-and-excitation enhancement module (DSEEM).

Results and discussion: To thoroughly assess the performance and generalization capability of MAF-Net, we conducted an extensive series of experiments on the clinical dataset of uterine fibroids from Quzhou Hospital of Traditional Chinese Medicine. Across all evaluation metrics, MAF-Net demonstrated superior performance compared to existing state-of-the-art segmentation techniques. Notably, it achieved Dice of 0.9126, Mcc of 0.9089, Jaccard of 0.8394, Accuracy of 0.9924 and Recall of 0.9016. Meanwhile, we also conducted experiments on the publicly available ISIC-2018 skin lesion segmentation dataset. Despite the domain difference, MAF-Net maintained strong performance, achieving Dice of 0.8624, Mcc of 0.8156, Jaccard of 0.7652, Accuracy of 0.9251 and Recall of 0.8304. Finally, we performed a comprehensive ablation study to quantify the individual contributions of each proposed module within the network. The results confirmed the effectiveness of the multi-receptive attention fusion module, the dual-path squeeze-and-excitation enhancement module, and the dual-scale attention enhancement module.