<![CDATA[Fetal cerebellum segmentation in ultrasound images is important for quantitative analysis of fetal brain development, yet it remains challenging due to speckle noise, low contrast, acoustic artifacts, and unstable anatomical boundaries. This study proposes an ROI-Based Shape-Prior Reconstruction method as a post-processing refinement stage for YOLOv8n-seg fetal cerebellum segmentation. A total of 294 fetal ultrasound images with manually annotated binary cerebellum masks were used and divided into training, validation, and testing subsets using a 70:20:10 ratio. YOLOv8n-seg generated the initial segmentation masks, while the proposed ROI-based reconstruction stage refined the foreground region using a convolutional autoencoder trained on ROI-based binary cerebellum masks. Compared with raw YOLOv8n-seg, the proposed method improved DSC from 0.9282 to 0.9302 and IoU from 0.8671 to 0.8708. Boundary performance also improved, with HD95 decreasing from 15.06 to 14.18 and ASSD decreasing from 5.38 to 5.20. Although these improvements were modest and not statistically significant, the proposed method produced smoother boundaries and more morphologically consistent segmentation outputs in the visual evaluation. These results indicate that ROI-based shape-prior reconstruction can serve as a lightweight refinement stage for improving boundary consistency in fetal cerebellum ultrasound segmentation. However, external validation with larger datasets is still required to assess generalization.]]>
