Attention-based recognition of coronary angiography projection views to reduce inter-view confusion

Background/Introduction Accurate recognition of standard coronary angiography projection views supports consistent interpretation and enables downstream quantitative analysis. In real-world data, adjacent projections can be visually similar (particularly cranial/caudal variants and near-neighbor oblique views), and class imbalance can amplify misclassification between neighboring views. Purpose To develop and evaluate a hybrid convolutional–transformer model for automatic classification of common coronary angiography projection views, with specific emphasis on reducing errors driven by inter-class visual ambiguity. Methods We trained a hybrid architecture that combines a residual convolutional encoder (ResNet-34) with a transformer attention decoder, enabling integration of global context with fine-grained spatial cues relevant to projection angle. The model was fine-tuned on 4,754 labelled angiography frames across six standard categories: AP caudal/LAO caudal, AP cranial/RAO cranial, LAO, LAO cranial, RAO, and RAO caudal. To address class imbalance and hard-to-separate examples, we used class-balanced focal loss, robustness-oriented augmentation (including blur and photometric perturbations), and a cosine-annealing learning-rate schedule. Model performance was assessed on an external validation set of 1500 held-out frames using overall accuracy, macro-averaged precision and recall, and class-level F1-scores. All reported results reflect performance on this external validation set. Results The model achieved 94.7% overall accuracy (95% CI: 93.5–95.8%), with macro-averaged precision of 94.5%, recall of 93.5%, and F1-score of 93.9%. Macro ROC-AUC was 0.962, indicating strong class separability despite visual similarity between neighboring projections. The highest-performing category was AP cranial/RAO cranial (F1-score 97.8%; recall 98.3%). The most challenging categories were AP caudal/LAO caudal and LAO cranial (F1-scores 91.7% and 91.3%, respectively), reflecting clinically confusable neighboring projections. Conclusion(s) A hybrid convolutional–transformer approach can deliver high-accuracy coronary angiography projection view classification while materially reducing confusions between visually adjacent projections. The key contribution is the use of transformer attention on top of convolutional features, paired with class-balanced focal optimization, to improve discrimination in clinically ambiguous views and support standardized angiographic analysis.