Semi-supervised Cardiac Cine-MRI Multi-Structure Segmentation via Temporal Consistency on Unlabeled Cine Frames
DOI: https://doi.org/10.62517/jike.202604410
Author(s)
Yali Wan1,*, Deyi Xiong2
Affiliation(s)
1School of Information Technology and Engineering, Guangzhou College of Commerce, Guangzhou, China
2Chongqing Academy of Metrology and Quality Inspection, Chongqing, China
*Corresponding author
Abstract
The Automated Cardiac Diagnosis Challenge (ACDC) supplies manual annotations for only two frames of each cine-MRI sequence (end-diastole and end-systole), although every sequence contains roughly 30 frames. We present TC-SS, a mean-teacher semi-supervised framework that puts the remaining unlabeled frames to work through three complementary objectives: temporal coherence between adjacent frames, agreement with the temporal mean prediction, and pseudo-label consistency between frames of similar appearance. The supervised branch keeps the compound boundary-aware loss of our previous work and is applied only to the annotated ED/ES frames. Under 3-fold cross-validation on ACDC, evaluated on the unified per-slice ED+ES protocol shared across the three technical papers of this series, TC-SS reaches a mean foreground Dice of 0.713 ± 0.011 against the supervised HybridUNet baseline (fold-0 mDice 0.863, an absolute gap of 0.150), while using 15 times as many frames per patient at no annotation cost. On a per-class basis TC-SS reaches RV Dice 0.669 ± 0.020, MYO Dice 0.686 ± 0.008 and LV Dice 0.784 ± 0.017, all below the supervised baseline (0.818, 0.866 and 0.905 for RV, MYO and LV respectively); the myocardium shows the most consistent gap across folds, which is consistent with the global temporal-consistency terms over-regularising the thin myocardial band. The widened gap relative to earlier patient-level ED-only reports reflects the unified evaluation on ED+ES slices: the ES frame is intrinsically harder than ED and pulls both numbers down, more so for TC-SS.
Keywords
Cardiac Cine-MRI; ACDC; Semi-Supervised Learning; Temporal Consistency; Mean Teacher
References
[1] Bernard O, Lalande A, Zotti C, et al. Deep Learning Techniques for Automatic MRI Cardiac Multi-structures Segmentation and Diagnosis: Is the Problem Solved? IEEE Transactions on Medical Imaging, 2018, 37(11):2514-2525. DOI:10.1109/TMI.2018.2837502.
[2] Tarvainen A, Valpola H. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in Neural Information Processing Systems (NeurIPS), 2017, 30:1195-1204.
[3] Sohn K, Berthelot D, Carlini N, et al. FixMatch: Simplifying Semi-Supervised Learning with Consistency and Confidence. Advances in Neural Information Processing Systems (NeurIPS), 2020, 33:596-608.
[4] Laine S, Aila T. Temporal Ensembling for Semi-Supervised Learning. International Conference on Learning Representations (ICLR), 2017.
[5] Hung W C, Tsai Y H, Liou Y T, Lin Y Y, Yang M H. Adversarial Learning for Semi-Supervised Semantic Segmentation. British Machine Vision Conference (BMVC), 2018.
[6] Li S, Zhang C, He X. Shape-aware Semi-supervised 3D Semantic Segmentation for Medical Images. Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2020, LNCS 12261:552-561. DOI: 10.1007/978-3-030-59710-8_54.
[7] Shen B, Huang X, Shi Y, et al. Dual Bidirectional Copy-Paste with Shape Constraint for Semi-Supervised Image Segmentation. IEEE International Symposium on Biomedical Imaging (ISBI), 2024:1-5. DOI:10.1109/ISBI56570.2024.10635399.
[8] Yu L, Wang S, Li X, Fu C W, Heng P A. Uncertainty-aware Self-ensembling Model for Semi-supervised 3D Left Atrium Segmentation. Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2019, LNCS 11765:605-613. DOI: 10.1007/978-3-030-32245-8_67.
[9] Bai Y, Chen D, Li Q, Shen W, Wang Y. Bidirectional Copy-Paste for Semi-Supervised Medical Image Segmentation. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023:11514-11524. DOI:10.1109/CVPR52729.2023.01108.
[10] Chen D, Bai Y, Shen W, Li Q, Yu L, Wang Y. MagicNet: Semi-Supervised Multi-Organ Segmentation via Magic-Cube Partition and Recovery. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023:23869-23878. DOI:10.1109/CVPR52729.2023.02286.
[11] Ma J, Li F, Wang B. U-Mamba: Enhancing Long-range Dependency for Biomedical Image Segmentation. arXiv: 2401.04722, 2024. DOI:10.48550/arXiv.2401.04722.
[12] Ronneberger O, Fischer P, Brox T. U-Net: Convolutional Networks for Biomedical Image Segmentation. Medical Image Computing and Computer-Assisted Intervention (MICCAI), 2015, LNCS 9351:234-241. DOI: 10.1007/978-3-319-24574-4_28.
[13] Oktay O, Schlemper J, Folgoc L L, et al. Attention U-Net: Learning Where to Look for the Pancreas. arXiv: 1804.03999, 2018. DOI:10.48550/arXiv.1804.03999.
[14] Isensee F, Jaeger P F, Kohl S A A, et al. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nature Methods, 2021, 18:203-211. DOI: 10.1038/s41592-020-01008-z.
[15] Chen J, Lu Y, Yu Q, et al. TransUNet: Transformers Make Strong Encoders for Medical Image Segmentation. arXiv: 2102.04306, 2021. DOI:10.48550/arXiv.2102.04306.
[16] Cao H, Wang Y, Chen J, et al. Swin-Unet: Unet-like Pure Transformer for Medical Image Segmentation. European Conference on Computer Vision Workshops (ECCVW), 2022, LNCS 13803:205-218. DOI: 10.1007/978-3-031-25066-8_9.
[17] He K, Chen X, Xie S, Li Y, Dollár P, Girshick R. Masked Autoencoders Are Scalable Vision Learners. IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022:16000-16009. DOI:10.1109/CVPR52688.2022.01553.
[18] Chen T, Kornblith S, Norouzi M, Hinton G. A Simple Framework for Contrastive Learning of Visual Representations. International Conference on Machine Learning (ICML), 2020, PMLR 119:1597-1607.