A Classification Model to Identify Target Selection Acceptability from a Library of Plans for Bladder Cancer Radiotherapy
Abstract
Purpose
We developed an algorithm to assess plan selection from a pre-existing plan library to unlock the full potential of plan-of-the-day (POTD) radiotherapy. The algorithm mimics nuanced human decision-making, automatically identifying the best plan (small- or medium-sized) based on the clinical target volume (CTV).
Methods
48 cone-beam CT (CBCT) images from 9 bladder cancer patients treated with POTD were used to develop and test the algorithm. Reflecting clinical practice, two radiographers achieved consensus for the best selected plan based on a visual assessment of CTV coverage to define soft-tissue registration and ground-truth acceptability.For each CBCT, the daily CTV and bladder was contoured by an oncologist. CTVintra was generated by anisotropically expanding the CTV to model anticipated bladder filling. Ten features describing spatial relationships between CTVintra and both PTVs were derived. These included scalar metrics (e.g. percent volumetric overlap) and metrics that preserved the anatomical orientation of discrepancies (e.g. 95th percentile of the maximum anterior non-overlapping distance). An L2-regularized logistic regression model with standard scaling was trained using leave-one-group-out cross-validation (9-fold). The coefficients associated with the ten metrics were estimated on the training set and applied to the held-out dataset to select a plan per fold.
Results
The acceptability accuracy (mean (std)) overall held-out test group was 0.919 (0.097). Model sensitivity to identify the correct acceptability was 0.947. The specificity to identify correct unacceptability was 0.917.
Conclusion
It is feasible to use a contour-based plan-selection algorithm that mimics the human plan selection process to automatically identify the best plan from a library. The tool accuracy is similar to the greatest reported concordance of online and offline review. In conjunction with autocontouring, not only may this tool improve uptake, concordance, and confidence in POTD protocols, it could be modified to assess plan acceptability in online adaptive workflows.