MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

An AI-Based Tumor Auto-Contouring Algorithm for Non-Invasive Intra-Fractional Tumor-Tracked Radiotherapy (nifteRT) On Linac-MR

J Yun1*, E Yip2, Z Gabos3, N Usmani4, D Yee5, K Wachowicz6, B Fallone7, (1) (2) (6) (7) Department of Medical Physics, Cross Cancer Institute, Edmonton, AB, CA, (3) (4) (5) Department of Radiation Oncology, Cross Cancer Institute, Edmonton, AB, CA

Presentations

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

Room: AAPM ePoster Library

Purpose:
To develop an AI-based lung, liver, and prostate tumor auto-contouring algorithm for nifteRT on linac-MR. NifteRT is a novel RT technique being developed to achieve < 3 mm geometric error in treating mobile tumors with radiotherapy.

Methods:
During nifteRT, our linac-MR will acquire continuous MR images of a mobile tumor. Accurate tumor auto-contouring in each MR image is critical, to allow for accurate tumor localization. We modified U-net for tumor segmentation from its surrounding anatomy. Using U-net, our algorithm can take (1) abstract, knowledge-based tumor shapes, and (2) pixel-by-pixel details of tumor/healthy tissue boundaries into account for segmentation. Hence, superior auto-contouring performance is expected even if the tumor resides in low-contrast areas with diffuse boundaries to its surroundings.

6 lung, 3 liver, and 3 prostate cancer patients were scanned with 3T MRI at 4 frames per second (2D steady-state free precession sequence, free breathing). For each patient, an expert delineated gold standard tumor contours (ROI_std) on 130 consecutive images. ROI_std were delineated in 3T images. However, to validate the algorithm in our 0.5T linac-MR environment, the 3T-acquired images were noise-degraded to reflect the worst-case scenario image quality characteristics of tumors at 0.5T.

For each patient, the first 30 ROI_std were used to train the algorithm. Once trained, the algorithm was applied to segment tumors for the remaining 100 pseudo-0.5T images, generating 100 ROI_auto. In each image, the Dice coefficient (DC), Hausdorff distance (HD), and centroid displacement (?d_centroid) were calculated between ROI_std and ROI_auto.

Results:
The algorithm achieved 87 – 96% DC, 0.9 – 3.6 mm ?d_centroid, and 2.4 – 10.4 mm HD from 12 patients' data. The algorithm successfully segmented tumors exhibiting substantial shape and location changes during breathing.

Conclusion:
An AI-based tumor auto-contouring algorithm for nifteRT was developed and validated with in-vivo MR images showing > 87% DC.

Funding Support, Disclosures, and Conflict of Interest: COI disclosure: Stockholder in MagnetTx Oncology Solutions

Keywords

Segmentation, MRI, Image-guided Therapy

Taxonomy

IM/TH- MRI in Radiation Therapy: MRI/Linear accelerator combined- IGRT and tracking

Contact Email