1336521 - Semi-supervised learning using labeled and unlabeled data for LGE CMR segmentation

Segmentation of the myocardium in late gadolinium enhancement (LGE) cardiac MR is critical to detect and quantify myocardial enhancement. While deep learning (DL) solutions show high accuracy, they depend strongly on the availability of large, diverse, annotated datasets. Annotating these images is a tedious, time-consuming task, requiring clinical expertise. It limits the size of datasets that can be used for training DL algorithms. We propose a simple semi-supervised (SS) method to exploit unlabeled data and compare its performance to fully supervised (FS) training with limited data.

Methods:

The model consists of a pair of UNets for segmentation [1]. To use unlabeled data, cross pseudo supervision [2], with differing weights initialization and data augmentation, was used. For analyzing the incremental effect of unlabeled data, experiments were done by varying the ratio of labeled and unlabeled data. The SS experiments used N% of data with ground truth (GT) labels (N = 3, 7, 12, 25, 50, 75) and the remaining (100-N)% with no GT. For comparison, a series of FS experiments were done, by training only on the corresponding labeled dataset.

LGE images acquired with either a 2D segmented inversion recovery (IR) gradient echo (GRE) or a single-shot IR balanced steady state free precession (bSSFP) pulse sequence were collected from 3 centers: Center 1: 539 patients/4934 images, Center 2: 34 patients/358 images (MAGNETOM Aera and Avanto, respectively, Siemens Healthcare, Erlangen, Germany) and Center 3: 15 patients/106 images from the publicly available EMIDEC challenge (1.5 & 3T Siemens Healthcare, Erlangen, Germany) [3]. To show improvements in accuracy & robustness, the model was trained and finetuned using 85% of patients from Center 1 and tested on the remaining data from all 3 centers.

Results: The details of the test sets and dice score (DSC) for each experiment are given in Table 1. The SS experiments outperformed the FS ones on both in-distribution (Center 1) and out-of-distribution data (Center 2 & 3). Using extra unlabeled data improved segmentation accuracy, as shown by significantly increased DSC and decreased standard deviation, compared to using only labeled data, except for 3 cases involving Center 3. While using only 3% of the labeled data, adding unlabeled data allows to achieve up to a 37-point gain in DSC. Additionally, by using only 50% of the labels, comparable results to the FS training are obtained. The results are illustrated in Fig. 1. Example contours are shown in Fig. 2. 

Conclusion:

Incorporating unlabeled data during training can help increase segmentation accuracy and robustness on diverse data, as shown by the superior results compared to FS training using a fraction of the available annotations. Even for limited amounts of annotated data, the performance was comparable to reported inter-expert variability [4].

The concepts presented are based on research results that are not commercially available. Future availability cannot be guaranteed.