TAT-Det: A Dataset for Testicular Appendage Torsion Detection and Segmentation in Ultrasound Images

Testicular appendage torsion (TAT) is the most common cause of acute scrotal pain in prepubertal children, yet its diagnosis remains challenging because of nonspecific clinical manifestations, subtle sonographic appearances, and the lack of standardized datasets dedicated to pediatric acute scrotal emergencies. To address this gap, we present TAT-Det, the first publicly available and expertly annotated ultrasound dataset specifically dedicated to TAT, and to the best of our knowledge, the first large-scale publicly available ultrasound dataset for pediatric acute scrotal emergencies. TAT-Det comprises 2,600 B-mode ultrasound images retrospectively collected from 1,203 patients at the Children's Hospital, Zhejiang University School of Medicine. Each image is accompanied by lesion-level bounding boxes for primary detection and localization, together with corresponding pixel-level segmentation masks for fine-grained delineation. The annotations were established through a rigorous multi-stage workflow involving independent annotation by board-certified pediatric sonographers, consensus review, and annotation refinement, providing reliable supervision for both localization and segmentation. To facilitate reproducible evaluation, we establish comprehensive baseline benchmarks for object detection and semantic segmentation using representative deep learning models. The benchmark results demonstrate the challenging nature of TAT localization and delineation while establishing reference performance for future algorithmic development. By combining a relatively large and clinically diverse collection of pediatric scrotal ultrasound images with expert-verified dual-level spatial annotations and standardized evaluation protocols, TAT-Det fills an important gap in publicly available resources for pediatric emergency ultrasound and provides a foundation for developing and systematically evaluating artificial intelligence methods for automated TAT analysis.