In this paper, the problem of orientation correction in cardiac MRI images is investigated and a framework for orientation recognition via deep neural networks is proposed. For multi-modality MRI, we introduce a transfer learning strategy to transfer our proposed model from single modality to multi-modality. We embed the proposed network into the orientation correction command-line tool, which can implement orientation correction on 2D DICOM and 3D NIFTI images.
-
truncation and concat:
- For the maximum pixel value of each 2D slice as
$X_{max}$ , truncation operations are performed on$X_t$ at the threshold of 60%, 80% and 100% of$X_{max}$ to obtain$X_{1t},X_{2t},X_{3t}$ . - Concatenated 3-channel image
$[X_{1t}, X_{2t}, X_{3t}]$ as$X'$
- For the maximum pixel value of each 2D slice as
- histogram equalization;
- random small-angle rotations, random crops (in training), and resize;
- z-score normalization
we pre-train the model on the bSSFP cine dataset, and then transfer the model to the late gadolinium enhancement (LGE) CMR or T2-weighted CMR dataset. On the new modality dataset, we first load the pre-trained model parameters. We freeze the network parameters of the backbone and retrain the fully connected layers on the new modality dataset.
MSCMR orient：链接: https://pan.baidu.com/s/1cE5i68YUNhXrzUpldTV6ow 提取码: mj6p
[1] Xiahai Zhuang: Multivariate mixture model for myocardial segmentation combining multi-source images. IEEE Transactions on Pattern Analysis and Machine Intelligence 41(12), 2933–2946, 2019
[2] Xiahai Zhuang: Multivariate mixture model for cardiac segmentation from multi-sequence MRI. MICCAI 2016, 581–588, Springer, 2016
[3] Ke Zhang and Xiahai Zhuang: Recognition and Standardization of Cardiac MRI Orientation via Multi-tasking Learning and Deep Neural Networks. MyoPS 2020, LNCS 12554, 167–176, Springer Nature, 2020, https://github.com/BWGZK/Orientation-Adjust-Tool