Comparison of Semi-supervised DL Methods for 3D Myocardial Segmentation in Whole-Heart Joint T1/T2 Mapping

Background: Myocardial tissue characterization with T1 and T2 mapping is key in detecting cardiovascular disease (CVD) ^[1-2]. Traditionally, 2D T1 and T2 maps are acquired sequentially under several breath-holds. Recently, 3D whole-heart joint T1/T2 maps from a single free-breathing acquisition have been demonstrated ^[3]. However, the large data from this single sequence increases segmentation and analysis time.
Deep learning has been effective in automatic 2D segmentations, but 3D applications remain limited due to the manual effort of labeling multi-contrast data. In this work, we compare three semi-supervised segmentation approaches for 3D left ventricle (LV) myocardial segmentation in whole-heart joint T1/T2 maps. We compare 1) a semi-supervised nnUNet^[4] that utilizes both labeled and unlabeled data for self-training, 2) a Students Discrepancy-Informed Correction Learning (SDCL) ^[5] teacher-student approach, and 3) a modified SDCL with Swin-UNet3D^[6].

Methods: The semi-supervised 3D nnUNet requires three steps: i) training the model with a small amount of labeled data, ii) use the trained model to infer pseudo labels from the unlabeled data and iii) train the model with the labeled and pseudo labeled data (Fig.1a). SDCL trains two student models in parallel and incorporates non-trainable discrepancies between them to guide self-correcting learning with the limited labeled data (Fig. 1b). SDCL^[5] originally used VNet and ResVNet. In our modified version, we propose replacing ResVNet with Swin-UNet3D^[6]. Data consisted of 3D T1/T2 maps with four interleaved volumes ^[3] from 50 subjects (10 healthy, 40 with suspected CVD), with 45 used for training under different labeled data combinations (N_labeled = 3, 5, 7, 9, or 11). Training was carried out for sufficient epochs for each method (100 for nnU-Net, 276 for SDCL). Segmentation was evaluated via Dice score (DSC) on five test subjects (unseen during training) and compared in terms of training and inference times.

Results: Results for all semi-supervised methods in comparison to manual segmentation and supervised nnUnet (N_labeled = 45) are shown in Fig.2, Fig.3 and Table 1. All methods achieve good performance, requiring only 11 labels to achieve similar results than the supervised nnUNet with 45 labels. Semi-supervised nnUNet showed a slight advantage over SDCL, with a DSC of 0.92 for nnUNet, of 0.90 for SCDL with ResVNet and of 0.90 for SCDL with Swin-UNet3D. Training time was 4 hours for nnUNet and 6 hours for SDCL. Inference times were 5s for nnUNet and 1 min for SDCL.

Conclusion: The investigated nnUNet and SDCL semi-supervised methods enabled accurate segmentation of 3D LV myocardium in T1 and T2 maps with 25% of the labelled data required for supervised nnUNet. Among the semi-supervised methods, nnUNet with pseudolabelling achieved slightly better performance in terms of metrics and inference time and SCL Swin-UNet3D improved over SCDL. Further evaluation and comparison in terms of generalizability is now warranted.

Table 1: Dice Score results for supervised nnUNet and semi-supervised nnUNet and SDCL approaches.

Supervised		Semi-supervised
Labeled	nnUNet	Labeled	Unlabeled	Semi-supervised nnUNet	SDCL (ResVNet)	SDCL (Swin-UNet)
3	0,850	3	42	0,852	0,863	0,860	DSC
5	0,885	5	40	0,894	0,882	0,880
7	0,892	7	38	0,896	0,877	0,889
9	0,906	9	36	0,910	0,894	0.909
11	0,908	11	34	0,923	0,898	0,902
45	0,900	-		-	-	-

Figure 1: Proposed frameworks for semi-supervised 3D myocardial segmentation. (a) Semi-supervised training workflow using nnUNet. An initial supervised training with a reduced set (N = 3, 5, 7, 9, or 11) of labeled images allows the model to generate pseudolabels for unlabeled data. These pseudolabels are incorporated into a second training stage, where they are assigned lower weights to account for their higher uncertainty compared to manual annotations. The resulting model is evaluated on a test set. (b) Students Discrepancy-informed Correction Learning (SDCL) framework. The SDCL adopts a Balanced Copy-Paste (BCP) strategy, merging labeled and unlabeled images to form mixed images. The teacher, updated via Exponential Moving Average (EMA) of Student A, generates pseudo-labels, which are combined with ground truths to produce mixed labels. Students A and B process the mixed images independently, and discrepancy and error masks are computed to guide correction learning.A modified version of this framework is also considered, where the second student network (ResVNet) is replaced with a Swin-UNet3D.
Figure2: Results for semi-supervised nnUnet and SDCL methods (Nlabeled = 11) in comparison to manual segmentation and supervised nnUnet (Nlabeled = 45) . Representative views are shown for T1 from a healthy subject.
Figure 3: Results for semi-supervised nnUnet and SDCL methods (Nlabeled = 11) in comparison to manual segmentation and supervised nnUnet (Nlabeled = 45) . Representative views are shown for T2 from a healthy subject.