Implementation of deep learning with convolutional block attention module for detecting collimator rotation errors in stereotactic radiosurgery quality assurance.

Stereotactic radiosurgery (SRS) requires sub-millimeter accuracy, making patient-specific quality assurance (QA) essential. Conventional Gamma Index analysis is often insufficient to detect subtle collimator errors. This study aimed to develop and evaluate a convolutional neural network (CNN) enhanced with a Convolutional Block Attention Module (CBAM) for automated error detection in SRS treatment plans. A total of 146 SRS plans were generated using the Eclipse v13.6 treatment planning system, including error-free and ± 1° collimator-rotated cases. Dose maps were extracted from DICOM RT Dose files, normalized, and augmented by flips, rotations, and Gaussian noise. Dose–volume histogram (DVH) features were also encoded. The CNN model with three convolutional blocks and CBAM was trained with stratified five-fold cross-validation and evaluated using accuracy, precision, recall, F1-score, and AUC-ROC. Independent validation was performed on 20 real clinical cases from the 108 Military Central Hospital. The model achieved an average accuracy of 68.2% (± 3.5%) and an AUC of 0.65 (± 0.04) across cross-validation folds. On the independent clinical dataset, the best model achieved 85% accuracy, with a recall of 0.90 and an F1-score of 0.86. The CNN was able to identify subtle collimator deviations that conventional Gamma analysis might miss. A CNN with CBAM demonstrated feasibility for detecting collimator errors directly from treatment planning system (TPS) dose distributions in SRS QA. Although preliminary and limited by dataset size and single-error type, this approach shows promise as a complementary tool to conventional QA methods, with potential to improve safety and efficiency of radiosurgery in resource-limited settings.