Five-parameter model reproduced physician-specific glioma CTV contouring patterns

KEY POINTS

• The retrospective single-institution study included 93 glioma patients treated by three radiation oncologists at Massachusetts General Hospital, including 87 glioblastomas, three astrocytomas, two anaplastic astrocytomas and one anaplastic oligodendroglioma.
• Instead of using a black-box segmentation network, the model represents CTV decisions with five interpretable parameters: overall GTV-to-CTV expansion, reduced expansion into the brainstem, reduced expansion into optic structures, and positional uncertainty allowances around the ventricles and falx.
• When individually fitted to physician contours, the model reproduced them closely across all three physicians, with median Dice similarity >0.95, 95% Hausdorff distance <5 mm, and surface Dice with 2-mm tolerance >0.83.
• In blinded clinical review, model-generated and physician-generated CTVs received the same overall average quality score of 2.3 on a 0-3 scale, where 3 was clinically acceptable and 2 required only minor edits. There was no statistically significant difference between model and clinical contours.
• The fitted parameters differed systematically between physicians, particularly at the brainstem and ventricles, demonstrating that some apparently subjective contouring differences could be represented quantitatively rather than simply labelled interobserver variability.
• A random-forest classifier using the five fitted parameters identified the physician's contouring style with 80% accuracy and a Cohen kappa of 0.68. Ventricular positional uncertainty was the most informative parameter, followed by overall GTV-to-CTV expansion.
• The model remains institution-specific and predominantly reflects glioblastoma practice. It does not incorporate performance status, molecular information or the full complexity of clinical decision-making, and multicentre prospective validation is still required.

CLINICAL TAKEAWAY

SOURCE

Radiotherapy and Oncology