Multi-center analysis of machine-learning predicted dose parameters in brachytherapy for cervical cancer

Background and purposeImage-guided adaptive brachytherapy (IGABT) is a key component in the treatment of cervical cancer, but the nature of the clinical workflow makes it vulnerable to suboptimal plans, as the theoretical optimal plan depends heavily on organ configuration. Patient anatomy-based quality-assurance (QA) with overlap volume histograms (OVHs) is a promising tool to detect such suboptimal plans, and in this analysis its suitability as a multi-institutional clinical QA tool is investigated.Materials and methodsA total of 223 plans of 145 patients treated in accordance with the current state-of-the-art IGABT protocols from UMC Utrecht (UMCU) and Erasmus MC (EMC) were included. Machine-learning models were trained to predict dose D2cm3 to bladder, rectum, sigmoid and small bowel with the help of OVHs. For this strategy, points are sampled on the organs-at-risk (OARs), and the distances of the sampled points to the target are computed and combined in a histogram. Machine-learning models can then be trained to predict dose-volume histograms (DVHs) for unseen data. Single-center model robustness to needle use and applicator type and multi-center model translatability were investigated. Performance of models was assessed by the difference between planned (clinical) and predicted D2cm3 values.ResultsIntra-validation of UMCU data demonstrated OVH model robustness to needle use and applicator type. The model trained on UMCU data was found to be robust within the same protocol on EMC data, for all investigated OARs. Mean squared error between planned and predicted D2cm3 values of OARs ranged between 0.13 and 0.40 Gy within the same protocol, indicating model translatability. For the former protocol cohort of Erasmus MC large deviations were found between the planned and predicted D2cm3 values, indicating plan deviation from protocol. Mean squared error for this cohort ranged from 0.84 to 4.71 Gy.ConclusionOVH-based models can provide a solid basis for multi-institutional QA when trained on a sufficiently strict protocol. Further research will quantify the model's impact as a QA tool.