A Machine-Learning Based Method for Inter-Institutional QA of MR-Based Brachytherapy Treatment Planning in Cervical Cancer

Abstract

PURPOSE/OBJECTIVE(S): Inter-institutional quality assurance (QA) of brachytherapy (BT) treatment planning is often based on expert judgment of a limited number of treatment plans. Cohort comparisons are of limited value as patient anatomy has a major impact on organs-at-risk (OAR) dose. Therefore, the aim of this study was to develop and test a QA tool that predicts OAR dose based on patient anatomy. MATERIALS/METHODS: 60 Patients (120 plans) from institute A (data A) and 14 patients (32 plans) from institute B (data B) were included, treated in accordance with EMBRACE II guidelines. Additionally, 71 MR-guided BT pre-EMBRACE II plans (71 patients) from institute B were included (data B'). Histograms of the overlap (OVHs) between delineated OARs and the high-risk CTV were used to objectify patient anatomy. Dimensionality of the OVH data was reduced by principal component analysis. A random-forest model was fitted to training OVHs and DVHs. Model performance was evaluated using leave-one-out cross-validation for data A. Then, different models were created and tested based on data splits according to institute (A versus B and A versus B'), applicator type (ovoid versus ring), application type (IC versus IC+IS). The models predict DVHs from OVHs, from which the D2cm3 of the OARs was computed. Model performance based on data A was evaluated by calculating the distribution (σ) of the difference between planned and predicted D2cm3 values (D2cm3, pl-pr), and the Pearson correlation coefficient (r) of these values. For the models based on the data splits it was tested if the D2cm3, pl-pr values fell within the 95%-confidence interval (CI) of the D2cm3, pl-pr values from data A. RESULTS: Leave-one-out validation of the model based on data A demonstrated predictability of the D2cm3 values for all OARs (bladder r = 0.64, rectum r = 0.75, sigmoid r = 0.88, small bowel r = 0.92). The distribution of D2cm3, pl-pr values was relatively constant for all OARs (bladder σ = 0.61 Gy, rectum σ = 0.56 Gy, sigmoid σ = 0.48 Gy, small bowel σ = 0.53 Gy). For the different data splits, models trained on one applicator or application type could predict D2cm3 values for the other applicator or application type within the CI. Training on data A and testing on data B resulted in predicted bladder D2cm3-values within the CI for 30/32 plans. In contrast, only 42/71 plans of data B' fit within the CI (Chi-squared test, P < 0.001). CONCLUSION: Our OVH-based model can predict D2cm3 values for all applicable OARs in a multi-center setting. The models are robust against differences in applicator and application type, and are sufficiently sensitive to distinguish differences in planning protocols. We believe that OVH-based QA can play an important role to assure treatment plan quality in multi-institutional studies.