Automatic CT Quality Assurance Artifact Detection Using Residual-Enhanced Teacher–Student Learning Artificial Intelligence
Abstract
Purpose
Routine CT quality assurance (QA) relies on visual inspection, where subtle non-uniformities can be difficult to identify consistently across scanners and readers. We developed an efficient, interpretable artificial intelligence-assisted tool that detects and localizes artifacts in CT QA images to reduce manual review burden.
Methods
We propose a CT QA artifact detection model (AutoCTAD), a teacher–student anomaly detector with residual-based multi-scale enhancement to amplify subtle non-uniformities while preserving smooth images. We randomly selected 16,734 normal studies (49,996 images) spanning 2023-01 to 2025-12 and curated 300 artifact images including ring, band, shading, tube-spit, and low-contrast smudge patterns. Normal images were defined by visual inspection and using a CT-number-based QA criterion. Artifact images were identified by physicist visual inspection and included cases with either passing or failing CT-number criteria, to capture clinically relevant artifacts that may not be reflected by CT number alone. Normal images were split into training/validation/testing sets of 39,837/9,959/200 images. An image-level decision threshold was set using a percentile criterion derived from train-normal scores. Pixel-level heatmaps were generated for localization and review. The model was trained and evaluated on a Linux system with a NVIDIA A100 GPU.
Results
On the test set (300 artifact, 200 normal), AutoCTAD achieved 97.8% accuracy, 98.0% precision, and 98.2% F1-score (295/300 artifacts detected; 194/200 normals correctly rejected). Mean end-to-end inference latency was 143 ms/image for 512×512 inputs on the NVIDIA A100, including residual-enhancement preprocessing. False negatives were subtle central smudges with acceptable CT number, whereas false positives were uniform phantom images with mild non-uniformity or smudge-like texture. Heatmaps localized suspicious regions consistent with visual findings and supported rapid adjudication.
Conclusion
AutoCTAD provides high-performance, interpretable artifact detection with localized heatmaps across diverse artifact types, while identifying failure modes for contrast-limited smudge pattern.