Radiation Physics

PV QA 3 - Poster Viewing Q&A 3

TU_14_3253 - The Impact of Dataset Size on an Artificial Intelligence-Guided Clinical Decision Support System for Radiation Therapy Planning

Tuesday, October 23
1:00 PM - 2:30 PM
Location: Innovation Hub, Exhibit Hall 3

The Impact of Dataset Size on an Artificial Intelligence-Guided Clinical Decision Support System for Radiation Therapy Planning
J. Chan1, J. Chen2, G. Valdes2, S. S. Yom2, A. Pattison3, C. Carpenter3, and T. D. Solberg1; 1University of California San Francisco, Department of Radiation Oncology, San Francisco, CA, 2University of California, San Francisco, San Francisco, CA, 3Siris Medical, Redwood City, CA

Purpose/Objective(s): Previously, we validated a machine learning-guided clinical decision support system for predicting achievable dosimetry for a given plan, based on target and organ-at-risk (OAR) feature similarity to prior plans stored in a searchable database. Here we investigate the impact of dataset size on the number of potentially achievable (matching) plans identified per patient for locoregionally advanced lung and oropharyngeal cancers and demonstrate that these plans are achievable.

Materials/Methods: 51 patients with locoregionally advanced lung cancer and 82 with oropharyngeal cancer who received curative-intent Intensity Modulated Radiation Therapy (IMRT) were used to build the search spaces. DICOM-RT datasets were processed using commercial plan-classification software and the number of matching plans was quantified in relation to database size. The lung dataset included dose to the lung, cord, esophagus, heart and tumor. The oropharyngeal dataset included dose to the oral cavity, larynx, esophagus, cord, brainstem and tumor. Matches for two lung patients were identified and loaded into a treatment planning system to demonstrate that matched plans can be achieved with the optimizer.

Results: We found that the number of matched plans for both datasets scaled linearly with the size of the search space. In the lung cohort, database sizes of 20, 30, 40, and 50 resulted in mean number of matches of 3.06 +/- 0.792, 3.47 +/- 0.912, 4.16 +/- 0.760, and 5.23 +/- 0.424, respectively (r = 0.979). In the oropharyngeal cohort, database sizes of 20, 40, 60, and 80 resulted in mean number of matches of 2.49 +/- 0.674, 4.08 +/- 0.939, 5.55 +/- 1.19, and 6.60 +/- 0.653 (r = 0.996). Representative comparisons of matched and achieved plans for two patients with lung cancer are shown in Table 1, demonstrating that comparable dosimetry from matched plans can be achieved during treatment planning.

Conclusion: We demonstrated the ability to identify an expected or better number of plan matches across varying sizes of search spaces. The number of potentially achievable matching plans increased modestly and linearly as the dataset size increased. These results have clinical implications for best methods to improve planning efficiency and enhance dosimetric quality metrics.
Patient 1 Patient 2
Matched Achieved Matched Achieved
PTV Max 71.8 71.2 73.1 72.4
PTV D99 64.6 65.5 60.1 61.3
PTV D95 66.4 66.3 64.4 64.2
Cord Max 34.2 21.4 41.1 39.2
Esophagus V50 13.6 4.0 0.0 0.0
Esophagus Max 62.4 61.2 45.3 47.0
Esophagus Mean 15.6 14.5 20.8 8.10
Heart Max 23.4 19.7 N/A N/A
Heart Mean 2.27 1.44 N/A N/A
Lung V20 25.1 26.0 18.3 19.0
Lung V5 57.3 49.0 41.8 42.0
Lung Max 72.1 71.0 67.8 71.7
Lung Mean 14.9 12.8 11.6 12.3
Table 1: Dosimetric comparisons for matched and achieved plans for two locoregionally advanced lung cancer patients.

Author Disclosure: J. Chan: None. J. Chen: Research Grant; Raysearch Laboratories, Accuray Incorporated, Siemens Medical Solutions. G. Valdes: None. S.S. Yom: Research Grant; Genentech, Merck, Bristol-Myers Squibb. royalty; UpToDate, Springer. Honoraria; ASTRO. Chair; Am Radium Society-Am College of Radiology. Treasurer; American Radium Society. A. Pattison: Partnership; Siris Medical, Inc. C. Carpenter: President and CEO; Siris Medical, Inc. T.D. Solberg: Speaker's Bureau; Brainlab. Partnership; Global Radiosurgery, LLC. Deputy Editor-In Chief; JACMP.

Jason Chan, MD

UCSF Radiation Oncology

Disclosure:
No relationships to disclose.

Presentation(s):

Send Email for Jason Chan


Assets

TU_14_3253 - The Impact of Dataset Size on an Artificial Intelligence-Guided Clinical Decision Support System for Radiation Therapy Planning



Attendees who have favorited this

Please enter your access key

The asset you are trying to access is locked. Please enter your access key to unlock.

Send Email for The Impact of Dataset Size on an Artificial Intelligence-Guided Clinical Decision Support System for Radiation Therapy Planning