Multi-sequence Cardiac MR Segmentation with Adversarial Domain Adaptation Network

Abstract

Automatic and accurate segmentation of the ventricles and myocardium from multi-sequence cardiac MRI (CMR) is crucial for the diagnosis and treatment management for patients suffering from myocardial infarction (MI). However, due to the existence of domain shift among different modalities of datasets, the performance of deep neural networks drops significantly when the training and testing datasets are distinct. In this paper, we propose an unsupervised domain alignment method to explicitly alleviate the domain shifts among different modalities of CMR sequences, e.g., bSSFP, LGE, and T2-weighted. Our segmentation network is attention U-Net with pyramid pooling module, where multi-level feature space and output space adversarial learning are proposed to transfer discriminative domain knowledge across different datasets. Moreover, we further introduce a group-wise feature recalibration module to enforce the fine-grained semantic-level feature alignment that matching features from different networks but with the same class label. We evaluate our method on the multi-sequence cardiac MR Segmentation Challenge 2019 datasets, which contain three different modalities of MRI sequences. Extensive experimental results show that the proposed methods can obtain significant segmentation improvements compared with the baseline models.

Deep learning‐based cardiac cine segmentation: Transfer learning application to 7T ultrahigh‐field MRI

“…In particular, the notion to provide data sequentially by individual may result in higher data requirements than necessary. There are various other routine cardiac MR examinations such as T 2 , T 1 , LGE, and even that require segmentation 41,42,45 . Transfer learning applications to image segmentation of such varying contrasts may benefit from the amount of annotated data and the framework provided in this study.…”

“…The state of generalization and model suitability for a specific data set can be monitored through sensitivity analysis. 39 In addition, it emphasizes why improvements in generalization [40][41][42] are needed and why we applied an additional step of transfer learning to 7T data.…”

A Deep Learning Based Cardiac Cine Segmentation Framework for Clinicians – Transfer Learning Application to 7T

“…With respect to the performance on 7T data this just means that, compared to the UKBB dataset, the Kaggle data set contains image patterns and characteristics more similar to the 7T data we acquired. In addition, it emphasizes why improvements in generalization [37][38][39] are needed and why we applied an additional step of transfer learning to 7T data. Due to differences in training data our initial models based on UKBB labels outperformed the UKBB model on the Kaggle data.…”

“…In particular, the notion to provide data patient by patient may result in higher data requirements than necessary. There are various other routine cardiac MR examinations such as T 2 , T 1 , LGE, and even T 2 * that require segmentation [38,39,42]. Transfer learning applications to image segmentation of such varying contrasts may benefit from the amount of annotated data and the framework provided in this study.…”

Adapt Everywhere: Unsupervised Adaptation of Point-Clouds and Entropy Minimization for Multi-Modal Cardiac Image Segmentation

“…The method transferred style, shape, and appearance from bSSFP images to LGE to generate synthetic samples, to train the segmentation network. Wang et al [31] proposed a fully end-to-end unsupervised method based on adversarial training to minimise discrepancies in both the feature and output space. Roth et al [32] couples classical methods of multi-atlas label fusion with deep learning by formulating noisy labels for unlabelled LGE images using the registration technique.…”