SS_CASE_UNet: An Attention-Enhanced Semi-Supervised Framework for Fetal Cerebellum Segmentation in Ultrasound Images

Accurate segmentation of the fetal cerebellum in ultrasound images is crucial for assessing fetal development and detecting prenatal abnormalities. However, this task remains challenging due to factors such as image noise, complex anatomical structures, and limited availability of annotated data, which is further compounded by the high cost and effort required for manual labeling. To address these challenges, we propose SS_CASE_UNet, a novel semi-supervised segmentation framework that enhances U-Net with attention mechanisms to better manage image noise and anatomical complexity. Additionally, a multi-stage semi-supervised training strategy effectively mitigates the scarcity of annotated data. The architecture integrates Squeeze-and-Excitation blocks for dynamic channel-wise feature recalibration and a Coordinate Attention block at the bottleneck to capture precise spatial and long-range dependencies. Our multi-stage training pipeline leverages both labeled and unlabeled data through iterative pseudo-label and re-training, improving generalization in low-annotation scenarios. Experimental results demonstrate that SS_CASE_UNet outperforms existing methods, achieving a Dice Similarity Coefficient (DSC) of 87.65%, along with high accuracy (99.08%), precision (93.49%), recall (82.34%), and Jaccard Similarity (81.78%). Despite incorporating advanced attention mechanisms, our model maintains a balanced complexity-performance trade-off. These results highlight SS_CASE_UNet as a robust and clinically practical solution for automated segmentation of the fetal cerebellum in ultrasound images.