J Yang¹, R Wang^1,2,3, Y Weng^2,3*, L Chen^2,3, Z Zhou⁴, (1) School of Artificial Intelligence, Xidian University, Xi'an, CN. (2) UT Southwestern Medical Center, Dallas, TX. (3) Medical Artificial Intelligence and Automation (MAIA) Lab, University of Texas Southwestern Medical Center, Dallas, TX. (4) University Of Central Missouri, Warrensburg, MO.

• 
  Y Weng

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: accurate substructure segmentation of brain gliomas has played an important role in tumor diagnosis, tumor characterization, and surgical planning. In general, the gliomas have three substructures: whole tumor (WT), tumor core (TC) and enhancing tumor (ET). In particular, we noticed that WT tissue include TC, and TC tissue include ET. Therefore, inspired by the multi-task deep learning approach, we aim to develop a hierarchical three dimensional U-Net (hU-Net) model which can exploit this prior information to improve the segmentation accuracy.


Methods: BraTS2018 dataset were used in this study. Dataset contained 285 patients, which include 75 low-grade gliomas and 210 high-grade gliomas. Every patient has four sequences: T1, T1Gd, T2, and, FLAIR. We select 228 patients as training set and the other 57 patients as the testing set. The proposed hU-Net model consists of three hierarchical cascade U-Net blocks. Each U-Net block was used to segment one substructure, and then the segmented mask was made a voxel-wise product with the original input image. The results were then used as the input of the next U-Net block. Three basic U-Net blocks were connected and the loss function was independently calculated. Finally, we utilized the multi-task deep learning paradigm to optimize the whole network.


Results: compared our hU-Net method with the baseline U-Net. For the proposed hU-Net, the average dice coefficient of all testing samples achieved 0.90, 0.86, and 0.74 for three substructures WT, TC and ET. Compare to baseline, our method obtained an improvement of 9.32%, 12.75%, and 3.15% for each category respectively.


Conclusion: proposed hU-Net is capable of modeling hierarchical structure of gliomas for brain tumor substructure segmentation. The quantitative and visual segmentation results demonstrated that the proposed method obtained better performance compared with the conventional methods described.

Funding Support, Disclosures, and Conflict of Interest: This work was supported in part by Engineering Research Centerof BigData Application in Private Health Medicine, Fujian ProvinceUniversity, Putian, Fujian(KF2020005).

Keywords

MRI, Segmentation

Taxonomy

IM- MRI : Machine learning, computer vision

Contact Email