Automatic Classification of Pre-Existing Plan Acceptability As a Decision-Support Tool to Streamline Online Adaptive Radiotherapy
Abstract
Purpose
Online adaptive radiotherapy (oART) is time-consuming and resource-intensive. To alleviate this, we developed a tool combining autocontouring and image registration to classify pre-existing treatment plans (Planspre), defined as the offline reference plan or a previous adapted plan, as acceptable or unacceptable for re-delivery given the daily magnetic resonance image (MRIdaily).
Methods
Data from 67 patients (328 MRIs) receiving radiotherapy for prostate cancer (40 Gy in 5 fractions) on an MR-linac were used. A two-stage deep-learning network was trained on 193 MRIs from 57 patients, with stage-1 segmenting the prostate and organs at risk (OARs), and stage-2 segmenting the seminal vesicles conditioned on stage-1 outputs. Pairwise registration between each Planpre and MRIdaily was performed by: (1) extracting the 40 Gy isodose surface from the dose distribution generated from Planpre online at a previous fraction, (2) converting this isodose into a mask, and (3) rigidly registering it to the CTV derived from the MRdaily autosegmentations. Plans were considered acceptable if predefined target and OAR dose-volume objectives were met. Tool-based dose-volume metrics were computed using autosegmentations and original dose distributions. Ground truth (GT) dose distributions (DoseGT) were obtained by recalculating each Planpre onto the corresponding MRIdaily, with dose-volume metrics evaluated using both autosegmentations (GTauto) and manual contours (GTmanual) to disentangle dosimetric from combined dosimetric–contouring effects.
Results
Evaluation on 125 MRIdaily–Planpre comparisons showed that 58 and 40 plans were acceptable according to GTauto and GTmanual, respectively, with 98% correctly identified as acceptable in both cases. Contouring variability increased false positives, with 7 and 25 unacceptable plans classified as acceptable for GTauto and GTmanual, respectively; however, associated deviations in dose–volume metrics were small (3% mean difference between tool and GTmanual).
Conclusion
This tool accurately identifies acceptable Planspre, even in the presence of contour variability, and can support decision-making to streamline oART.