KEY POINTS
- This retrospective study included 498 clinical helical tomotherapy plans from two institutions: 286 TomoH plans and 212 RadixAct plans. Treatment sites included brain, thorax, liver, and prostate, with fraction doses ranging from 1.8 to 6 Gy.
- Measurement-based patient-specific quality assurance used ArcCHECK at both institutions. Binary outcomes were defined using global gamma passing-rate thresholds of 95% for 3%/2 mm and 90% for 2%/2 mm, with a 10% low-dose cutoff.
- Investigators extracted 72 plan-complexity features and 851 three-dimensional dose-distribution radiomic features. Support vector machine classifiers were developed using complexity features alone, dose radiomics alone, or both combined, with institution-specific 80% training and 20% independent test sets.
- The hybrid model achieved the highest test-set discrimination in every setting. AUCs were 0.774 and 0.938 for 3%/2 mm in Institutions 1 and 2, respectively, and 0.820 and 0.825 for 2%/2 mm.
- At 100% sensitivity for the 3%/2-mm criterion, the hybrid model achieved specificities of approximately 33% and 50% in the two institutions. This means some passing plans could potentially be screened out from measurement while retaining all threshold-failing plans in those test sets.
- Dose radiomics dominated the selected features: 37–44 of the 50 retained predictors were dose-distribution features, compared with 6–13 plan-complexity metrics. Wavelet-derived texture descriptors were selected more frequently than features extracted from the original dose distribution.
- The Institution 2 AUC of 0.9375 for 3%/2 mm remained above chance in an exact permutation test (p<0.0139). However, performance declined markedly when models were transferred between institutions, indicating sensitivity to planning systems, delivery platforms, commissioning, detector workflows, and local data distributions.
- Most plans had high gamma passing rates, producing substantial class imbalance and limited specificity. The model cannot detect output drift, multileaf collimator malfunction, or mechanical instability and therefore cannot replace machine QA or measurement-based verification without prospective external validation.
CLINICAL TAKEAWAY
Combining delivery complexity and spatial dose information appears more informative than either feature domain alone for virtual tomotherapy QA. For now, the approach should be regarded as a locally trained triage tool, not a generalizable substitute for measurement-based PSQA.