Joint segmentation and landmark detection for cardiac MRI with multi‐scale inter‐task feature fusion and gradient‐bridged anatomical constraints
Haiyang Chen, Zhihao Liu, Chenxi HuAbstract
Background
Segmentation and landmark detection are essential yet complex tasks in Cardiac MRI (CMR) for quantitative assessment of the heart. Existing automatic CMR analysis methods typically handle the two tasks independently, overlooking their synergies that could improve the performance of both tasks.
Purpose
To develop an accurate and efficient multi‐task learning (MTL) framework that jointly performs segmentation and landmark detection for CMR.
Methods
The proposed SLaDNet‐GaB framework integrates a Segmentation and Landmark Detection Network (SLaDNet) with a Gradient‐as‐Bridge (GaB) loss. SLaDNet adopts a shared feature encoder and two separate decoders to generate segmentation and landmark predictions, respectively. Inter‐Task Feature Fusion (ITFF) modules are inserted between the decoders at multiple resolution scales to facilitate bidirectional information exchange via channel‐wise attention. Along with two intra‐task losses, the dedicated GaB loss is further introduced during training to enforce inter‐task anatomical consistency. Specifically, the loss guides the predicted landmarks toward the segmentation objective boundaries using gradient maps. We comprehensively evaluated the proposed method against both single‐ and MTL baselines on two public CMR datasets: a cine dataset (ACDC; 105/15/30 subjects for training/validation/testing) and a T1 mapping dataset (STONE; 150/20/40 subjects for training/validation/testing).
Results
SLaDNet‐Gab outperformed the baselines in both segmentation (dice scores: 92.37/94.27% on ACDC/STONE, p < 0.05) and landmark detection (point‐to‐point distances: 2.22/1.51 voxels on ACDC/STONE, p < 0.05). For downstream tasks, the proposed method consistently achieved lower strain and T1 quantification errors than the other methods in most myocardial segments, with per‐segment measurements closely agreeing with the manually derived values. Moreover, SLaDNet‐GaB demonstrated relatively low training times (11.4/3.2 hours on ACDC/STONE) and an inference time of 6.12 msec./sample suitable for clinical adoption.
Conclusions
The proposed multi‐tasking method achieves improved accuracy in CMR segmentation, landmark detection, and downstream quantification, while maintaining high computational efficiency.