Cardiac Segmentation and Spatiotemporal Modelling

ST-Net

ST-Net studies cardiac cine MRI segmentation using temporal information when segmentation labels are limited to selected cardiac phases.

Role

Co-developer · Third author

Period

Mar 2026 – Aug 2026

Research areas

Medical imaging · Temporal modelling

Submitted · AISTATS 2027

Overview

Support cardiac image analysis with limited annotated frames.

How it works

ST-Net applies a shared two-dimensional encoder and decoder across a cardiac cine sequence. A Temporal Attention Gate (TAG) at the bottleneck uses temporal depthwise convolution, channel mixing and time-dependent weights for residual recalibration. Supervision is limited to end-diastolic (ED) and end-systolic (ES) frames; cyclic area and centroid regularisation supports consistency across the sequence.

**Temporal recalibration with sparse labels**
Shared 2D encoding and decoding surround the TAG bottleneck. ED/ES-only supervision and cyclic area/centroid regularisation relate segmentation to the full cine sequence.
AI-assisted method illustration

Selected contributions

Developed ST-Net for segmentation across cardiac cine MRI sequences.
Compared model performance on 50 ACDC cine MRI cases.

Research status

Third author of “ST-Net: Temporal Recalibration for Cardiac Cine Segmentation under ED/ES-only Labels”; submitted to AISTATS 2027.

Search website