MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Automatic Localization and Segmentation of the Pancreas in Motion Artifact-Free CBCT Reconstructions Using Fully Convolutional Networks

P Jordan*, A Wang , J Star-Lack , J Van Heteren , Varian Medical Systems, Palo Alto, CA

Presentations

(Tuesday, 7/31/2018) 9:30 AM - 10:00 AM

Room: Exhibit Hall | Forum 5

Purpose: To demonstrate feasibility of automatic localization and segmentation of the pancreas and other abdominal organs in CT scans as well as CBCT scans free from motion artifacts.

Methods: A fully convolutional network (FCN) model using a modified V-Net architecture was developed to automatically localize and segment the pancreas in abdominal CT and CBCT scans. The training dataset consisted of 72 contrast enhanced abdominal 3D CT scans and 72 simulated breath-hold CBCT scans. Breath-hold CBCT scans were simulated from CT using an in-house simulation and FDK reconstruction software. Simulations were performed in half-fan beam geometry and included treatment couch insertion, forward scatter, beam hardening, electronic noise, and quantum noise. Accuracy of model-generated contours in terms of mean surface distance (MSD) and the Dice coefficient was evaluated on a validation dataset containing 10 CT scans and 10 breath-hold CBCT scans with manual contours.

Results: Despite pancreatic volumes ranging from 55.2 to 98.2 cm³ within the validation dataset, good segmentation accuracy was achieved on CT scans (MSD: 1.2 ± 0.3 mm, Dice: 0.85 ± 0.03) and breath-hold CBCT scans (MSD: 1.3 ± 0.4 mm, Dice: 0.84 ± 0.03). Model inference took less than 100 ms using GPU hardware.

Conclusion: The current work shows that FCN models are a promising approach for fast and accurate automatic localization and delineation of abdominal soft tissue structures in CT scans as well as CBCT scans free from motion artifacts. Similar models may be employed in many areas of radiotherapy including treatment planning and CBCT-guided online adaptive radiotherapy. Future efforts will examine model accuracy on clinical CBCT scans acquired using breath-hold techniques.

Funding Support, Disclosures, and Conflict of Interest: Authors are employees of Varian Medical Systems.

Keywords

Segmentation, Cone-beam CT, Image Guidance

Taxonomy

IM- Cone Beam CT: Machine learning, computer vision

Contact Email