Click here to


Are you sure ?

Yes, do it No, cancel

Simultaneous Multi-Organ Segmentation Using Multi-Source Adaptive MR Fusion

L Zhang1*, F Yin2 , J Cai3 , (1) Duke University Medical Center, Durham, NC, (2) Duke University Medical Center, Durham, NC, (3) The Hong Kong Polytechnic University, Kowloon, Yau Tsim Mong


(Monday, 7/30/2018) 9:30 AM - 10:00 AM

Room: Exhibit Hall | Forum 2

Purpose: To develop a simultaneous multi-organ segmentation method using multi-source adaptive MR fusion.

Methods: T1-weighted (T1-w), T2-weighted (T2-w), T2/T1-weighted (T2/T1-w), and diffusion weighted (DWI) MR images were acquired for eight liver cancer patients under GE 1.5T MR scanner. Four sets of digital human phantoms (XCAT) (T1-w, T2-w, T2/T1-w and DWI) were generated using average of the eight patient MRI signals and noises for each organ. These four MRI sets were fed into our in-house developed multi-source adaptive MR fusion program, from which 625 new contrast MRI sets were generated. Each voxel is vectorized by image intensity of these new images. Unsupervised learning was used to group the voxels using k-means clustering method with k ranging from 2 to 16. To evaluate the segmentation quality, organ edge was detected using Sobel filter for source MRI and auto-segmented MRI sets. Union of all source MRI edges is used as reference. Normalized cross-correlation (NCC) of reference to auto-segmented MRI edge, as well as to each source MRI edge were assessed.

Results: Multiple organs (liver, lung, heart, kidney, spleen, muscle, vertebral body, blood vessel and tumor) were automatically segmented on a PC workstation within 2 minutes, which is substantially faster than manual contouring and can be further accelerated on a GPU. More organs were segmented with an increasing cluster number. For cluster number 4 and above, the auto-segmented image edge has higher NCC to the reference than any of the source MRI. The NCC peaked at 0.962 when 8 cluster was used for the segmentation.

Conclusion: A novel simultaneous multi-organ segmentation method based on multi-source adaptive MR fusion was developed and its feasibility has been demonstrated in XCAT digital phantoms with real patient image intensities. Different from existing machine learning based methods, the proposed method has great potential to achieve auto-segmentation without cohort learning.

Funding Support, Disclosures, and Conflict of Interest: This project was funded by NIH grant R21CA165384.


Image Fusion, MRI, Segmentation


IM/TH- image segmentation: MRI

Contact Email