Clinical Experience with AI-Assisted Gross Tumor Volume Contouring In the Setting of Stereotactic Radiosurgery for Brain Metastases
Abstract
Purpose
A growing number of patients with increasingly many brain metastases (BMs) are being treated with stereotactic radiosurgery. Artificial intelligence-assisted gross tumor volumes (AIGTVs) can aid the radiation oncologist in contouring the gross tumor volumes (GTVs). We implemented AIGTVs for prospective use in the clinical environment, and conducted a study to evaluate the long-term potential of this solution.
Methods
Informed by a failure mode and effects analysis, treatment planners and oncologists followed a structured workflow to create and review AIGTV contours in the treatment planning system. An existing 3D V-Net deep learning model was applied to coregistered CT and MR images to produce the AIGTVs. 141 cases comprising 130 patients were contoured and treated by four oncologists. Each AIGTV was classified as true or false positive, then modified or deleted to prepare the final volumes. Clinical cases were monitored for process failures related to AIGTVs. Data on sensitivity and false positives were collected for intact BMs, and true positives were assessed for segmentation quality.
Results
A total of 1418 BMs were correctly identified. Patients were treated to a median of 5 BMs (IQR 2-11). 1206 BMs were treated, of which 1064 (88.2%) were detected as AIGTVs. There were 321 false positives (2.3/case) and 137 false negatives (1.0/case) . True positive AIGTVs going to treatment had a median Dice coefficient of 0.91 (IQR 0.85-0.97), and a median Hausdorff distance of 1.4 mm (IQR 1.0-1.9). The median edited volume as a percentage of GTV volume was 17% (IQR 6-30). No adverse events related to AIGTVs were observed.
Conclusion
We found that our solution is safe and accurate for assisting the oncologist in the management of complex cases with multiple BMs. Future efforts will address improving ease of use and the BM model performance.