QA of DL-based dose calculations: determination of the application range

Abstract

Purpose or Objective Deep learning (DL)-based dose engines are powerful and fast methods for secondary dose checking, however their inputs should be constantly monitored to ensure compatibility with the intended application range. This abstract explores anatomical checks as QA on the applicability of our DL-based dose engine on various patient anatomies. Materials and Methods Our network was developed using the DeepDose framework and was trained on the dose of individual multi-leaf-collimator (MLC) segments. The training data included abdominal tumour segments (oligometastatic, rectal, prostate) with varying size and shape, treated on a 1.5 T MRI radiotherapy system. Due to the fact that our trained network was already robust enough to geometrical variations, a set of descriptive features was extracted from the training data in order to encode further anatomical characteristics of the patient anatomies. Having calculated the 3D projection of each segment through the patient and couch, the following features were extracted: radiological depth at the isocenter plane, density at the isocenter plane and average density, all within the 3D projection. The accepted range per feature was set to a 5-to-95 percentile range, thus defining a working envelope space. Then, the segments of seven unseen tumour cases (lung, cervical, pancreatic, endometrial) were classified by comparing their inputs against the envelope. A segment was flagged when any of its features lay outside the corresponding range. Additionally, a DL-based dose calculation was performed for these segments. The evaluation of our method was performed per segment using the envelope classification of the new tumour anatomies and their agreement with the ground truth dose following a 3%/3mm gamma analysis and a pass rate threshold of 95%. Results The range of operation defined from the features of the training set is presented in Table 1. The additional patient segments used for evaluation were categorized based on their agreement with the envelope as accepted or flagged cases. A sensitivity of 0.94 and specificity of 0.78 were reported. An overview of the correctly flagged test segments (true positives) is presented on Figure 1, alongside with the corresponding features responsible for raising the flag. Conclusion We presented a method that defines the application range of our DL-based dose engine, by analyzing anatomical features of a large variety of abdominal tumour patients. We used this analysis to perform a fast QA check on arbitrary MLC segments: any segments lying within the envelope can be safely handled by the dose engine, while segments located outside its boundaries are flagged by our algorithm for additional checking. Our next steps will focus on improving the specificity and complexity of our method as well as adding more features in our envelope analysis. In the near future we aim to expand the tumour sites handled by our dose engine in order to introduce it in a clinical setup for secondary dose calculations.