Abstract
Fetal cardiac structure localization requires features that follow the size and orientation of individual anatomical targets. We present Geometry-guided Regional Interaction and Transformation for Oriented Bounding Box Detection (GRIT-OBB), an oriented method that uses corrected proposal geometry to guide feature extraction and bounding-box refinement. Starting from horizontal proposals, a supervised correction stage estimates oriented bounding boxes, which define the sampling regions for rotated position-sensitive alignment. Then, a geometry-guided relation residual combines regional features using Gaussian bounding box representations and Wasserstein distances, followed by classification and refinement. Finally, rotated intersection-over-union losses directly supervise both proposal correction and localization. We evaluated GRIT-OBB against six methods on three datasets: FOCUS, OCD, and AU-OPG, comprising 300, 120, and 901 images, respectively. On the FOCUS dataset, GRIT-OBB achieved a mean AP50 of 47.74%, AP75 of 8.53%, and mAP50:95 of 16.81%, corresponding to improvements of 4.74%, 1.8%, and 2.01% over the strongest baseline, respectively. On the OCD dataset, GRIT-OBB achieved a mean AP50 of 73.14%, AP75 of 18.97%, and mAP50:95 of 30.77%, corresponding to improvements of 12.40%, 13.54%, and 11.28% over the strongest baseline. On the AU-OPG dataset, GRIT-OBB achieved a mean AP50 of 82.65%, AP75 of 59.66%, and mAP50:95 of 50.66%, corresponding to an improvement of 2.41%, 7.27%, and 3.88% compared with the strongest baseline. In an ablation study, removing supervised geometric correction reduced performance on the FOCUS dataset from 47.74% to 43.44% for AP50, while removing both IoU losses decreased performance from 8.53% to 6.59% for AP75. These results support our hypothesis that geometry-guided feature alignment and direct supervision at successive localization stages can improve object detection for fetal cardiac structures.