A Single Institution Evaluation of a Machine Learning Based Decision Support Software for Head and Neck Radiotherapy
L Zhuang1*, E Bowers1, M Pankuch2, M Posner1, (1) Northwestern Lake Forest Hospital, Lake Forest, IL, (2)Northwestern Medicine Chicago Proton Center, Warrenville, IL
Presentations
PO-GePV-T-257 (Sunday, 7/25/2021) [Eastern Time (GMT-4)]
Purpose: To evaluate the performance of a machine learning (ML) based decision support software (InsightRT, Siris Medical, CA) for head and neck (H&N) radiation therapy.
Methods: Twenty-two sets of H&N patients’ data were retrospectively evaluated in this study. All patients were treated with volumetric modulated arc therapy and planned using the Monaco treatment planning system (Elekta Inc.). Three simultaneous integrated boost (SIB) treatment schemes were prescribed including PTV60/54-59.3/53.9 (11 patients), PTV69.96/59.4-63/54-59.3/53.9 (5 patients) and PTV66/59.4-63/54-59.3/53.9 (6 patients). The H&N prediction model of InsightRT was built upon a total of 813 patients with the aforementioned three Rx schemes. D95 for all PTVs (PTV69.96, PTV66, PTV60, PTV<59.3) and organ at risk (OAR) constraints including cord max (CM), mandible max (MM), left parotid mean (LPM) and right parotid mean (RPM) were compared between planned and predicted data through a paired t-test. Software running time was also evaluated to assess the efficiency of the clinical workflow.
Results: Statistically, there was no significant difference between planned value and predicted value for D95 of PTVs (pPTV69.96=0.45, pPTV66=0.3, pPTV60=0.07, pPTV<59.3=0.38) and OAR constraints (pCM=0.49, pMM=0.09, pLPM=0.46, pRPM= 0.28). Prediction could be performed in less than 5 seconds for each individual case, and structures could be automatically associated provided the structure labeling followed pre-defined rules, thus demonstrating a robust algorithm. In addition, matched patient data used for prediction were provided to assist dosimetric tradeoff decisions during the planning process.
Conclusion: InsightRT could provide plan result prediction in good agreement with clinically approved plans for H&N radiotherapy in a rapidly produced and automated fashion. Predicted plan results and matched patient data could assist with dosimetric tradeoff decisions during the planning process and help prevent unnecessary plan modifications or replanning cycles.
ePosters
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email