DCLA-Net: dual-encoder cross-level alignment network for ultrasound-based ectopic pregnancy mass segmentation

Benefiting from its real-time imaging capability and non-invasive nature, ultrasound imaging has become an important modality for the diagnosis of ectopic pregnancy. However, the segmentation of ectopic pregnancy masses in ultrasound images remains a challenging task due to low image contrast, indistinct boundaries, severe noise, and substantial variations in lesion size and shape. To tackle these issues, we propose DCLA-Net, a dual-encoder cross-level alignment network for ectopic pregnancy mass segmentation.

Specifically, DCLA-Net adopts a dual-encoder architecture composed of a convolutional neural network (CNN) branch and a Transformer branch, which are designed to capture local details and global contextual information. To enhance the integration of features from different semantic levels, a multi-scale context-aware fusion module (MSCAFM) is introduced to strengthen feature interaction and alignment across hierarchical representations. In addition, cross-level connections are established between the encoder and decoder to promote information propagation and alleviate the loss of fine details during feature transmission. Furthermore, a dual-attention fusion module (DAFM) is incorporated into the decoding stage to adaptively integrate channel-wise and spatial features, thereby improving feature representation capability.

To evaluate the effectiveness of the proposed method, extensive experiments were conducted on a clinical Ectopic Pregnancy Mass dataset and a publicly available Carotid dataset. On the Ectopic Pregnancy Mass dataset, DCLA-Net achieved Dice of 0.8455, Mcc of 0.8425, and Jaccard of 0.7341. On the Carotid dataset, it obtained Dice of 0.9597, Mcc of 0.9588, and Jaccard of 0.9227. Compared with state-of-the-art methods, DCLA-Net consistently achieves superior performance across all evaluation metrics on both datasets. In addition, ablation studies were performed to further validate the effectiveness of each proposed component. The results demonstrate that the dual-encoder architecture, MSCAFM, and DAFM all contribute positively to the overall segmentation performance.