Click here to

Session: Machine Intelligence Efficacy and Quality I [Return to Session]

Prospective Clinical Experiences of Using the First Pediatric Deep-Learning Auto-Contouring models for Cranio-Spinal Irradiation (CSI)

S Zieminski*, S M MacDonald, T I Yock, Y Wang, Department of Radiation Oncology, Massachusetts General Hospital, Harvard Medical School, Boston, MA

Presentations

MO-A-BRC-4 (Monday, 7/11/2022) 7:30 AM - 8:30 AM [Eastern Time (GMT-4)]

Ballroom C

Purpose: To report the prospective clinical experiences of using the first ever deep-learning auto-contouring models for pediatric radiotherapy. These models were developed to automate one of the most time-consuming sites for contouring – cranio-spinal irradiation (CSI).

Methods: Two deep-learning models were trained by patients aged 4 to 19, one to auto-contour thecal sac, each kidney and lung, and the other for whole vertebrae. They were tested on retrospective patients and commissioned for clinical use. The models were prospectively applied to ten children aged 5 to 11, six adolescents aged 12 to 17, two young adults aged 18 and 19, and two adults. The auto contour (AC) was reviewed and edited (if needed) by a dosimetrist before being sent to physician review. Each AC was compared to the approved planning contour (PC), quantitatively using Dice similarity coefficient (DSC) and mean distance to agreement (MDA); and qualitatively using a four-point scoring system: 1 (no editing or minor editing in 1 min), 2 (moderate editing in 3 min), 3 (significant editing with efficiency gain) and 4 (rejected). The approximate dosimetrist editing time for each AC was compared to the estimated time for manual-contouring, calculated by multiplying the drawing time/slice with the number of slices for thecal sac, vertebrae and each kidney, and as the whole organ for each lung.

Results: The mean DSC was >0.96 for thecal sac, >0.97 for kidneys, >0.98 for vertebrae and nearly 1 for lungs. The mean MDA was <0.1 mm for lungs and <0.5 mm for the rest. The mean score was 1 for lungs and <1.15 for the rest. On average, 93 minutes were saved for thecal sac, 55 for vertebra, 4 for kidneys and 2 for lungs.

Conclusion: Our prospective results indicated the high-quality ACs led to substantial time saving of ~2.5 hours for the dosimetrist.

Keywords

Image Processing, Computer Vision

Taxonomy

IM/TH- Image Segmentation Techniques: Machine Learning

Contact Email

Share: