Answers are grounded in this paper and cite their sources.
Accurate segmentation of uterine fibroids in ultrasound images is of great importance for computer-aided diagnosis and clinical assessment. However, this task remains difficult because ultrasound images often suffer from speckle noise, weak tissue contrast, indistinct lesion boundaries, and considerable variations in fibroid morphology. To improve segmentation accuracy under these challenging conditions, this study proposes DSCA-Net, a deep semantic and channel-spatial attention network for uterine fibroid segmentation. Specifically, the proposed network follows an encoder-decoder framework and incorporates several dedicated feature enhancement components. Firstly, a deep semantic attention gate module (DSAGM) is introduced into the deeper encoding stages to strengthen high-level semantic representation and guide the network to focus on fibroid-related regions. Secondly, a channel-spatial attention refinement module (CSARM) is embedded in the skip pathways to selectively refine encoder features before feature fusion, thereby reducing the interference of irrelevant background information. In the decoder, a residual feature refinement module (RFRM) is designed to enhance the recovery of spatial details and improve the fusion of multi-level features. In addition, a weighted efficient channel attention module (WECAM) is applied near the prediction layer to further recalibrate informative channel responses and stabilize feature activation. Experimental results on the uterine fibroid ultrasound dataset collected from Quzhou Hospital of Traditional Chinese Medicine, the public brain tumor dataset, and the digital database thyroid image (DDTI) dataset demonstrate the effectiveness and generalization capability of DSCA-Net across different medical imaging tasks. On the uterine fibroid ultrasound dataset, DSCA-Net achieved Dice, Mcc, and Jaccard scores of 0.9174, 0.9140, and 0.8479. On the public brain tumor dataset, the corresponding scores were 0.8473, 0.8464, and 0.7509, while on the DDTI dataset, DSCA-Net obtained Dice, Mcc, and Jaccard values of 0.7767, 0.7417, and 0.6371. In addition, comprehensive ablation experiments were conducted to verify the effectiveness of each key component in the proposed network. Computational efficiency was also evaluated by comparing model parameter counts and inference times with those of representative segmentation methods. DSCA-Net contains 15.07 million parameters and requires 163 ms per inference step under the reported experimental setting. These results demonstrate the segmentation performance of DSCA-Net across three medical imaging datasets and provide a quantitative assessment of its computational cost.
No takes yet. Share an insight, caveat, or question.
Zhou et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: