Automatic Detection of Multiple Artifacts in Computed Tomography Images by Self-Attention Guided Vision Transformer

Automated detection of image artifacts is crucial for enhancing the efficiency and quality of computed tomography (CT) workflows. This study presents a vision transformer (ViT)-based model for the simultaneous classification and localization of multiple artifacts in CT images across diverse body regions. We introduce a self-attention-guided class activation mechanism that aggregates selective gradients to generate precise artifact activation maps (AAMs), which visually highlight affected regions. To improve localization fidelity, register tokens are incorporated into the model, sharpening its focus on salient artifact features. On an independent test set, the model achieved outstanding performance, with an accuracy of 0.9947, sensitivity of 0.9968, specificity of 0.9984, and an F1 score of 0.9950. The generated AAMs provide radiologic technologists with interpretable visual guidance, bridging the gap between automated detection and clinical verification. This work establishes a unified, interpretable deep-learning framework for CT artifact analysis and presents a novel design paradigm for quality assurance in medical imaging.