MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

A Novel CNN-Based Adversarial Method with Markov Random Field Enhancement for Prostate and Organs-At-Risk Segmentation in Pelvic CT Images

Z Zhang1*, Z Zhang2, T Zhao3, B Sun4, (1) Washington Univ. In St. Louis, Computer Science and Engineering,Saint Louis, MO;(2) Washington Univ. In St. Louis, Computer Science and Engineering, Saint Louis, MO; (3) Washington University School of Medicine, St. Louis, MO; (4) Washington University in St. Louis, St. Louis, MO

Presentations

(Thursday, 7/16/2020) 10:30 AM - 12:30 PM [Eastern Time (GMT-4)]

Room: Track 4

Purpose: We develop a novel CNN-based adversarial deep learning method to automate the multi-organ segmentation of CT images on pelvic CT images in an end-to-end and one forward-propagation manner.

Methods: Planning CT and contours for 110 patients with prostate cancer were retrospectively selected and divided into sets for training, validation, and testing with a train/validation/test ratio of 90/10/10. The proposed adversarial multi-residual multi-scale pooling Markov Random Field (MRF) enhanced network (ARPM-net) implements an adversarial training scheme. A segmentation network and a discriminator network were trained jointly, and only the segmentation network is used for prediction. The segmentation network integrates a newly designed Markov Random Field block into a revised version of multi-residual U-net. The discriminator takes the product of the original CT and the prediction/ground-truth as input and classifies the input into fake/real. Multi-scale pooling layers are introduced to reduce memory usage and increase spatial resolution during pooling. An adaptive loss function is adopted to enhance the training on small or low contrast organs. The accuracy of modeled contours was measured with the Dice similarity coefficient (DSC) and Hausdorff distance (HD) using clinical contours as a reference to the ground truth. The accuracy of the model was compared to other popular deep learning methods.

Results: The proposed ARPM-net outperformed several existing deep learning frameworks and MRF methods and achieved state-of-the-art performance on the test dataset. The average DSC on the prostate, bladder, rectum, left-femur, and right-femur are 0.88(±0.05), 0.97(±0.03), 0.87(±0.05), 0.97(±0.01), and 0.97(±0.01), respectively. The average HD(mm) on these organs are 1.45(±1.86), 1.88(±1.84), 2.23(±1.17), 1.97(±1.60), 1.88(±1.33).

Conclusion: The proposed ARPM-net method is designed for the automatic segmentation of tumors and organs-at-risk in pelvic CT images. With adversarial fine-tuning, ARPM-net provides state-of-the-arts accurate contouring of multiple organs on CT images and has the potential to facilitate routine pelvic cancer delineation process.


Funding Support, Disclosures, and Conflict of Interest: This research is supported by Varian Research Grant.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email