Purpose: To evaluate the quality and efficiency of prostate treatment plan optimization using Oncospace dose predictions
Methods: The Oncospace web application predicts achievable DVHs based on novel shape relationship features, including overlap volume histograms (OVHs) and projection OVHs, using a random forest model trained on a large dataset of prior prostate plans. After obtaining predictions for 40 recently treated plans (“TxP”; 70 Gy in 28 fractions), new Oncospace plans (“OsP”) were generated by 3 consecutive optimizations according to typical dosimetrist workflows: (1) initial optimization using dose predictions (OAR weights set to 1); (2) dynamic adaptation of weights based on optimization objective values; and (3) addition of hot/cold spot ROIs. Comparison plans were created using the Pinnacle Autoplanning tool (“AP”), with similar OAR goals as Oncospace plans but with standard doses (no predictions). DVH metrics for all three plans were compared using the Friedman test, with a significance level of 0.002 accounting for multiple tests.
Results: OsP and AP were completed in 14.1±4.6 and 26.9±9.4 minutes on average, respectively. With OsP and AP normalized for PTV V[100%]=98%, no significant difference in PTV maximum dose was observed. AP demonstrated significantly better dose conformity and spilloff, but conformity was within clinical limits for all patients in all trials, and dose spilloff was within clinical limits for 97.5%, 97.5%, and 100% of TxP, OsP, and AP, respectively. DVH metrics were not statistically different for rectum V[57.45Gy], V[66.3Gy], and D[0.035cc]. OsP achieved significantly lower rectum V[35.35Gy], but compared to TxP, OsP had higher bladder V[35.35Gy], V[57.45Gy], V[66.3Gy], and D[0.035cc].
Conclusion: Plans optimized with Oncospace predictions can be generated in less time than Autoplans and with similar quality. Compared with original treatment plans, dose predictions uncovered tradeoffs that reduced rectum V[35.35Gy] and femoral head dose at the expense of slightly higher but clinically acceptable bladder coverage.
Funding Support, Disclosures, and Conflict of Interest: Oncospace, Inc. has received NSF SBIR grant funding, a portion of which has supported the current work.
Optimization, Treatment Planning
TH- External Beam- Photons: Treatment planning using machine learning/Knowledge Based Planning/automation