Automatic Triggering of Online Adaptive Radiotherapy for Liver Cancer via Integration of Multi-Dimensional CT Features
Purpose/Objective(s):
Significant interfractional anatomical variations (e.g., liver deformation, respiratory motion, gastrointestinal filling) during liver cancer radiotherapy compromise dose delivery accuracy. Although online adaptive radiotherapy (oART) can correct such errors, its clinical use is hindered by subjective and labor-intensive decision-making. This study aims to develop an artificial intelligence framework to automatically trigger oART for liver cancer patients by integrating morphological, radiomic, and dosimetric features.
Materials/Methods:
We retrospectively analyzed 419 daily fan-beam CT (FBCT) fractions from 24 liver cancer patients treated with curative radiotherapy. The planned dose was recalculated on each FBCT. Two radiation oncologists defined the ground-truth oART trigger labels based on morphological changes of targets and organs-at-risk on FBCT, combined with dosimetric criteria (e.g., PTV D95% < 95% prescribed dose, liver V30 > 33%, stomach Dmax > 50 Gy). Using 371 fractions from 21 patients, a support vector machine (SVM) classifier was developed based on CT-derived radiomic and dosimetric features, evaluated via repeated five-fold cross-validation. Feature selection with least absolute shrinkage and selection operator(LASSO) identified a stable set of discriminative features, predominantly shape-based and wavelet-derived texture features. The model was validated on an independent set of 54 fractions from 3 patients, with predictions compared against the physicians¡¯ labels.
Results:
Across 20 repetitions of five-fold cross-validation, the model achieved a mean training Area Under the Curve (AUC) of 0.91 and a mean validation AUC of 0.81, with a corresponding mean validation accuracy of 0.74. The best-performing cross-validation run yielded a validation AUC of 0.85, accuracy of 0.78, sensitivity of 0.78, and specificity of 0.78. Receiver operating characteristic (ROC) curves demonstrated consistent discriminative performance across folds, with individual fold AUCs ranging from 0.83 to 0.92, indicating good model stability and generalization. These results suggest that CT radiomic features combined with an SVM classifier provide robust performance for outcome discrimination in this cohort. In the independent test set, the weighted model outperformed physician consensus (accuracy: 0.76 vs 0.74,recall:0.82 vs 0.78), effectively reducing the risk of missed triggers.
Conclusion:
The proposed AI decision tool can accurately identify fractions requiring adaptation in liver cancer radiotherapy, potentially reducing clinical workload and standardizing the triggering process to promote the wider application of oART in liver cancer.
