2778

Using deep domain adaptation method to improve liver segmentation performance in Limited MRI images
Meng Dou1,2, Ailian Liu3, Yu Yao1,2, ZheBin Chen1,2, Han Wen1,2, Xu Luo1,2, and Ying Zhao3
1Chengdu Institute of Computer Application, Chinese Academy of Sciences, Chengdu, China, 2University of Chinese Academy of Sciences, Beijing, China, 3Department of Radiology, the First Affiliated Hospital of Dalian Medical University, Dalian, China

Synopsis

In clinical practice, it is too expensive to collect large-scale labeled Magnetic Resonance Imaging liver scans which constrains the segmentation performance. However, we noted that plenty of labeled Computed Tomography datasets aimed at liver have been published. Inspired by this, we proposed a deep domain adaptation method which can exploit the published CT datasets to improve the segmentation performance on MRI images. Our experiments showed that the liver segmentation performance is boosted on limited labeled MRI images (20 cases). Lastly, our method achieved competitive performance on both modality images. This work will be benefit to computer-aided diagnosis and treatment planning.

Purpose

To accurate and automatically segment liver based a small amount of labeled MRI images, we proposed a deep domain adaptation method to exploit the published CT data, which can reduce the labeling workload significantly.

Materials and Methods

This research retrospectively collected 20 cases’ MRI images from Department of Radiology, the First Affiliated Hospital of Dalian Medical University and 130 cases’ CT images from LiTS dataset. These contrast enhanced MRI images including: arterial, portal venous and delayed, which were all derived using Functool software on GE AW4.6 workstation and one radiologist with 10 years’ experience reviewed the data and manually delineated the region of interests (ROIs) at each slice of three modalities images.Before training procedure, all images are firstly resized to $$$128*128*64$$$ . As shown in Figure 1, the CT and MRI images input into deep model in a training iteration. Then encoder path constructed by 3-demension convolution kernel and down sample operation is employed to extract high level feature. After that, we use the upsample operation and convolution to construct decoder path which can reconstruct the feature map to original size. The skip connection is also employed to preserve contextual information from the encoder counterpart for decoder path. Finally, the deep model output a probability map which indicate the prediction class of each pixel. The performance of the deep model is evaluated by Volumetric dice and Average Symmetric Surface distance using the surface-distance package6.

Results

As shown in Table 1, we list the Volumetric dice (VD) and Average Symmetric Surface distance (ASSD) of two models on different domain. Comparing along the columns, we observe that 3D U-net for CT achieve a better performance than its counterparts on MRI images which may benefit from larger data volume. Interestingly, there has been a slightly decline in terms of volumetric dice when we train the 3d U-net on multi-modality data. We Speculate that the distinct physical principles of the underlying image acquisition give rise to this drop which also indicate that 3D U-net can’t exploit the cross-modality efficiently. Compared to the above method, our method further boosts the segmentation performance in terms of volumetric dice and Average Symmetric Surface distance.

Discussion

Our method can efficiently employ the labeled CT data to improve the liver segmentation performance in MRI images. This method can also enlarge data volume in some extent. However, one limitation of this research lies in data preprocessing. Before training, all images are resized to $$$128*128*64$$$ , which ignore the anisotropy and lost some details. In future work, we will employ the resample strategy to improve the segmentation performance further. Additionally, we will also employ this method to other medical image segmentation task and evaluate its’ performance.

Conclusion

We transfer the deep domain adaptation method (3D DAU-net) into liver segmentation task in MRI images. By employing open-source labeled CT data, the performance in MRI images get enhanced. We believe this work will be benefit to computer-aided diagnosis and treatment planning.

Acknowledgements

This research was supported by the National Natural Science Foundation of China under Gaint 6197010131.

References

  1. Heimann T, Van Ginneken B, Styner M A, et al. Comparison and evaluation of methods for liver segmentation from CT datasets[J]. IEEE transactions on medical imaging, 2009, 28(8): 1251-1265.
  2. Wang M, Deng W. Deep visual domain adaptation: A survey[J]. Neurocomputing, 2018, 312: 135-153.
  3. Bilic P, Christ P F, Vorontsov E, et al. The liver tumor segmentation benchmark (lits)[J]. arXiv preprint arXiv:1901.04056, 2019.
  4. Huang C, Han H, Yao Q, et al. 3D U^2-Net: A 3D Universal U-Net for Multi-domain Medical Image Segmentation[C]//International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, Cham, 2019: 291-299.
  5. Çiçek Ö, Abdulkadir A, Lienkamp S S, et al. 3D U-Net: learning dense volumetric segmentation from sparse annotation[C]//International conference on medical image computing and computer-assisted intervention. Springer, Cham, 2016: 424-432.
  6. https://github.com/deepmind/surface-distance
  7. Chollet F. Xception: Deep learning with depthwise separable convolutions[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2017: 1251-1258.

Figures

Note:Volumetric dice:(VD), Average Symmetric Surface distance (ASSD), all metrics are computed by surface distance package6. 3D U-net follow the practice in [5]. Multi-modality means two datasets (CT and MRI) are integrated into one dataset.

Figure 1. Overview of 3D DAU-net. 3D DAU-net follow the practice in 3D U-net5 include three components: (1) encoder path; (2) decoder path; (3) skip connection. The encoder and decoder paths both contain four levels at different resolutions. Within each level, depthwise separable convolution block is applied as the basic block4 and is used to replace the standard 3×3×3 convolution block. The skip connection is also employed to preserve contextual information from the encoder counterpart for decoder path.

Figure 2. Deep separable convolution block. In a training batch, CT and MRI data are sampled equally. Then using separable convolution7 to extract feature. Here, each domain has its own channel-wise kernel. Next a shared pointwise convolution layer is employed to extract common feature.

Figure 3. Visualization of different models’ prediction. A: 3D U-net for CT and MRI respectively. B: 3D U-net for Multi-modality.

Proc. Intl. Soc. Mag. Reson. Med. 29 (2021)
2778