T Wang*, Y Lei, J Roper, J Shelton, T Liu, X Yang, Emory University, Atlanta, GA

• 
  T Wang

• 
  
• 
  
• 
  
• 
  
• 
  
• 
  
• 
  

PO-GePV-M-333 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: Female pelvic anatomy is challenging for conventional auto-segmentation algorithms due to the large variations of tumors – both size and location – that results in a broad, diverse range of anatomical abnormalities. This work aims to develop an automatic method for the segmentation of multi organ female pelvic anatomy on CT images based on deep learning.

Methods: Our proposed method, named hierarchical target activation network, consists of four subnetworks, i.e., feature extractor, fully convolutional one-state object detector (FCOS), hierarchical block and mask module. The feature extractor is used to extract informative features from CT images. The FCOS is used to locate the volume-of-interest (VOIs) of multiple organs. The hierarchical block is used to enhance the feature contrast around an organ boundary and improve the accuracy of organ classification. The mask module then segments each organ from the refined feature map within the VOIs. We conducted a three-fold cross-validation study using 45 gynecologic cancer patient cases. The CT images and manual contours of bowel, bladder and rectum were used as training targets and the ground truth in the evaluation. The Dice similarity coefficient (DSC), 95th percentile Hausdorff distance (HD95), center-of-mass distance (CMD) and volume difference (VD) were calculated between an automatic and manually drawn contour to quantify segmentation accuracy.

Results: The proposed method achieved average DSC values of approximately 0.9 for all organs. Performance was best for the bladder. Both the bladder and rectum segmentations agreed with the ground truth to within 10 cc. On average, the HD95 and CMD were less than 10 mm and 4 mm, respectively, for all organs.

Conclusion: Quantitative results demonstrated the feasibility of the hierarchical target activation network-based segmentation method. The proposed method has the potential for clinical implementation in the current radiation therapy workflow for GYN cancer patients.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email