Room: AAPM ePoster Library
To develop a machine-learning (ML) model for brain V12Gy/V60% prediction of LINAC-based Single-Iso-Multiple-Targets (SIMT) SRS and study the longitudinal effects of the model’s application in clinic.
This ML model analyzes a SIMT case using the following statistics: 1)dose prescription; 2)target number and volumes; 3)modified center-of-mass distances; 4)equivalent spherical surface areas. These statistics are used by a Gradient Boosted Trees regression model to predict V12Gy/V60%.
Four versions of the model were trained. Model(v1) was built using 71 SIMT plans; Model(v1-Alpha) was built using 27 of the 71 plans planned by the same planner (Alpha). Both models were implemented in a clinical treatment planning system as a one-click scripting-based GUI execution. During the following 3-month study period, the V12Gy/V60% predictions were provided to the planners as guidelines. Seventeen recent plans (8 by Alpha) accrued in this period were added to the 71-plan cohort via replacing 17 oldest plans (8 by Alpha). Two new versions, Model(v2) and Model(v2-Alpha), were trained using the updated 71/27-plan cohort. For independent model tests, 19 cases (8 by Alpha) were used, where 7 were accrued in the study period (3 by Alpha). V12Gy/V60% mean prediction errors (MPEs) of the 4 model versions were compared.
Model(v1)/Model(v1-Alpha) MPEs of the 19/8 test cases were 1.67cc/0.90cc, and the corresponding MPEs of Model(v2)/Model(v2-Alpha) were 1.55cc/0.59cc. These results suggest an improved modelling accuracy from v1 to v2 after 3-month clinical implementation. In the evaluation of 7/3 recent test cases, Model(v1)/Model(v1-Alpha) MPEs were 1.51cc/1.52cc, and the corresponding MPEs of Model(v2)/Model(v2-Alpha) were reduced to 1.22cc/0.90cc. These results imply the improvements of both inter-planner(v2 vs.v1) and intra-planner(v2-alpha vs.v1-alpha) SIMT planning quality consistency.
An ML model for SIMT SRS V12Gy/V60% prediction was successfully developed. The presented longitudinal study suggests the great value of the model’s application in SIMT planning quality consistency improvement.