A Nazib*, K Boehm, V Paroder, S Shah, Y Lakhman, H Veeraraghavan, Memorial Sloan Kettering Cancer Center, New York, NY

• 
  A Nazib

• 
  
• 
  
• 
  
• 
  
• 
  
• 
  
• 
  

PO-GePV-M-306 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: Develop a deep learning (DL) segmentation of adnexal ovarian tumors to facilitate consistent automated radiomics and tumor response monitoring analyses.

Methods: A multiple resolution residual deep segmentation network (MRRN) was trained as a generational adversarial network (segmentor as generator) and an adversarial discriminator that used joint distribution of CT images and the segmentation to regularize the network training. The discriminator used Wasserstein distance metric to better distinguish the distribution of CT densities within segmented tumors to differentiate algorithm and expert segmentations. The network was further improved with texture distribution losses by combining the discriminator with a VGG19 to extract texture features to represent a vectorial distribution of textures and CT densities. Segmentation accuracy was evaluated against standard loss metrics such as cross entropy and Dice losses. Networks were trained with identical training and testing sets of contrast enhanced CT scans with 4-fold cross-validation from 460 patients with ovarian cancers.

Results: MRRN trained with Wasserstein distance using VGG textures or WGAN-VGG) was the most accurate with a tumor detection rate of 0.71 and median Dice similarity coefficient (DSC) of 0.68. It was similarly accurate as MRRN-GAN as well as MRRN-WGAN. It was significantly more accurate than the same network trained with cross-entropy and Dice losses (p < 0.001).

Conclusion: Incorporating texture information within tumors using an adversarial discriminator network improved the detection and segmentation accuracy of a network compared to standard losses.

Funding Support, Disclosures, and Conflict of Interest: Sohrab Shah is a consultant and shareholder of Canexia Health Inc. Yulia Lakhman Consultant, Calyx AI. K.M.B. is supported by NCI award F30CA257414 and a NIGMS MSTP grant T32GM007739 to the Weill Cornell/Rockefeller/Sloan Kettering Tri-Institutional MD-PhD Program.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email