Generating 3D images of VMAT plans for predictive models and activation maps associated with plan deliverability

Abstract

Background: Intensity modulation with dynamic multi-leaf collimator (MLC) and monitor unit (MU) changes across control points (CPs) characterizes volumetric modulated arc therapy (VMAT). The increased uncertainty in plan deliverability required patient-specific quality assurance (PSQA), which remained inefficient upon Quality Assurance (QA) failure. To prevent waste before QA, plan complexity metrics (PCMs) and machine learning models with the metrics were generated, which were lack of providing CP-specific information upon QA failures.

Purpose: By generating 3D images from digital imaging and comminications in medicine in radiation therapy (DICOM RT) plan, we proposed a predictive model that can estimate the deliverability of VMAT plans and visualize CP-specific regions associated with plan deliverability.

Methods: The patient cohort consisted of 259 and 190 cases for left- and right-breast VMAT treatments, which were split into 235 and 166 cases for training and 24 cases from each treatment for testing the networks. Three-channel 3D images generated from DICOM RT plans were fed into a DenseNet-based deep learning network. To reflect VMAT plan complexity as an image, the first two channels described MLC and MU variations between two consecutive CPs, while the last channel assigned the beam field size. The network output was defined as binary classified PSQA results, indicating deliverability. The predictive performance was assessed by accuracy, sensitivity, specificity, F1-score, and area under the curve (AUC). The gradient-weighted class activation map (Grad-CAM) highlighted the regions of CPs in VMAT plans associated with deliverability, compared against PCMs by Spearman correlation.

Results: The DenseNet-based predictive model yielded AUCs of 92.2% and 93.8%, F1-scores of 97.0% and 93.8% and accuracies of 95.8% and 91.7% for the left- and right-breast VMAT cases. Additionally, the specificity of 87.5% for both cases indicated that the predictive model accurately detected QA failing cases. The activation maps significantly differentiated QA failing-labeled from passing-labeled classes for the non-deliverable cases. The PCM with the highest correlation to the Grad-CAM varied from patient cases, implying that plan deliverability would be considered patient-specific.

Conclusion: This work demonstrated that the deep learning-based network based on visualization of dynamic VMAT plan information successfully predicted plan deliverability, which also provided control-point specific planning parameter information associated with plan deliverability in a patient-specific manner.