YOLOv13-Seg: Rethinking Instance-Aware Panoptic Segmentation for Volumetric Medical Imaging

Yue YaoOpenReview 2026 · OpenReview

Abstract

Efficiently parsing volumetric medical images into semantically meaningful regions and distinct instances is crucial for quantitative diagnosis, yet remains a formidable challenge. Current leading approaches often rely on heavy-weight 3D architectures with prohibitive computational costs, or treat semantic and instance segmentation separately, failing to capture their intrinsic synergy. In this paper, we rethink panoptic segmentation from an efficient detection-first perspective. We introduce YOLOv13-Seg, a novel, unified framework that extends the highly efficient YOLOv13 architecture for joint semantic and instance-aware segmentation in volumetric data. Drawing inspiration from Hypergraph-based Adaptive Correlation Enhancement (HyperACE) and Full-Pipeline Aggregation-and-Distribution (FullPAD), our core innovations include a 3D Panoptic Detection Head that concurrently predicts voxel-wise semantics and instance embeddings, and a lightweight Axial-Spatial Context Fusion module that models long-range dependencies. This design enables efficient capture of complex spatial relationships and seamless feature routing across scales. Extensive experiments on major benchmarks (BraTS, MSD) demonstrate that YOLOv13-Seg achieves state-of-the-art performance. It not only surpasses strong baselines like nnUNet in Panoptic Quality (PQ) and Dice score, but also shows a clear performance and efficiency advantage over its predecessors, YOLO12-N and YOLO11-N, in the same framework. Ultimately, our model delivers a 5.7× faster inference speed than previous 3D panoptic methods, making real-time volumetric analysis clinically viable.
BibTeX
@misc{yao2026yolov,
  author = {Yue Yao},
  title = {{YOLOv13-Seg: Rethinking Instance-Aware Panoptic Segmentation for Volumetric Medical Imaging}},
  year = {2026},
  url = {https://openreview.net/forum?id=ZjY5HaoIm9},
  howpublished = {OpenReview},
}

Loading related content...

Figure