MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel
  1. Home
  2. Meeting Program
  3. All
  4. Auto-Segmentation of Pelvic OARs On MRI Multi-Sequence Using An Fused-Unet

Auto-Segmentation of Pelvic OARs On MRI Multi-Sequence Using An Fused-Unet

Zesen Cheng1,2, Tianyu Zeng1,2, Yimei Liu1, Lijuan Lai2, Xin Yang1*, Sijuan Huang1, (1) Sun Yat-sen University Cancer Center; State Key Laboratory of Oncology in South China; Collaborative Innovation Center for Cancer Medicine; Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangzhou, Guangdong, 510060, China, (2) Department of Electronic and Information Engineering, South China University of Technology, Guangzhou, Guangdong, 510641, China

Presentations

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

Room: AAPM ePoster Library

Purpose: the increase application of MR-sim in RT, OARs auto-segmentation on MRI multi-sequence is becoming more and more important.

Methods: commonly MRI sequences (T1, T1DIXONC, and T2) and five classic pelvic OARs, including Bladder/Rectum/Anal Canal/Femoral(L)/Femoral(R), were used in this study.
Firstly, the traditional Unet was achieved as the baseline, which only used T1 images with an Adam-optimizer. The initial learning rate was 0.0001, The batch size was set to 4 and the maximum number of training epochs was 45. The model file with the best training performance within 45 epochs was selected for testing. Secondly, the Unet could be divided into three stages (input/middle/output). And the fused muti-sequence information could be introduced into above three stages. Three Fused-Unet models could be established (FUnet-input, FUnet-middle, FUnet-output), respectively. Thirdly, tri-/bi-sequence combinations (T1+T1DIXONC+T2; T1+T1DIXONC, T1+T2 and T1DIXONC+T2) were further explored for the FUnet-middle performance. 82 patients, were delineated by a radiation oncologist and verified independently by another one. 75 patients were randomly chosen as training set with the weighted cross entropy as the loss function. And the remaining 7 cases were regarded as test set. Dice similarity coefficient (DSC) was used to evaluated FUnet segmentation performance.

Results: took about 15 seconds to segment one patient's MR data automatically using FUnet, which is greatly improves the efficiency compared with manual delineation. All the OARs mean DSC was 0.894. And FUnet-middle had the highest DSC of 0.907±0.019, with Bladder 0.900±0.083/Rectum 0.823±0.121/Anal Canal 0.735±0.071/Femoral(L) 0.932±0.012/Femoral(R) 0.897±0.037, respectively. Furthermore, multi-sequence (tri-/bi- sequence) with FUnet-middle were also better than a single T1-sequence.

Conclusion: pelvic OARs of MRI multi-sequence could be auto-segmented rapidly with an accuracy result, and the FUnet has the potential to integrated more sequences in MRI images contouring.

Keywords

MRI, Segmentation, Contour Extraction

Taxonomy

IM/TH- image Segmentation: MRI

Contact Email

Additional Links