MICCAI 2026 paper debuts PCC embedding for single-slice cardiac MRI orientation
One CMR slice now yields LV orientation with 3.39° accuracy and 86% mIoU
Cardiac magnetic resonance (CMR) imaging relies on accurate left ventricular (LV) orientation to produce reliable short-axis views for downstream analysis. Existing approaches treat orientation as a discrete classification problem or require multiple intersecting slices, making them brittle when faced with arbitrary slices. A new MICCAI 2026 paper introduces a fundamentally different setting: jointly localizing the LV and estimating its continuous 3D orientation from a single CMR slice.
To solve this, the team proposes the Polar-Coupled Circular (PCC) embedding, a continuous and unambiguous orientation representation that avoids the discontinuities of traditional Euler angles or quaternions. They also built a scalable benchmark by automatically sampling slices from volumetric CMR segmentation datasets. Experiments across four datasets show strong generalization, with an average mIoU of 86.18% and an average angle deviation of just 3.39°. This work establishes a new task for single-slice LV modeling and offers a geometry-consistent framework for spatially informed CMR analysis. Code is publicly available.
- PCC embedding jointly localizes LV and estimates 3D orientation from one arbitrary CMR slice
- Achieves 86.18% average mIoU and 3.39° average angle deviation across four datasets
- Accepted to MICCAI 2026; code available on arXiv for reproducibility
Why It Matters
Single-slice LV orientation estimation removes the need for multi-slice scans, enabling faster, more robust cardiac analysis in clinical workflows.