Automated screening of computed tomography using weakly supervised anomaly detection.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Computer Vision

Background:

• Current AI for CT screening relies on supervised learning with heavy annotation needs or anomaly detection with lower performance.
• This study addresses the limitations of existing AI methods in CT screening by proposing a novel approach.

Purpose of the Study:

• To develop and evaluate a weakly supervised anomaly detection (WSAD) algorithm for CT screening.
• To reduce the annotation workload compared to traditional supervised learning methods.
• To improve the performance of anomaly detection in CT scans.

Main Methods:

• A novel WSAD algorithm was developed, trained using scan-wise normal and anomalous annotations.
• The algorithm utilizes an AR-Net-based convolutional network with dynamic multiple-instance learning loss and center loss.
• Publicly available datasets (RSNA brain hemorrhage and COVID-CT) were retrospectively analyzed.

Main Results:

• The WSAD algorithm successfully predicted anomaly scores for CT slices without slice-wise annotations.
• The brain CT dataset achieved a slice-level area under the curve (AUC) of 0.89, with sensitivity, specificity, and accuracy of 0.85, 0.78, and 0.79, respectively.
• Annotation requirements were reduced by 97.1% compared to slice-level supervised learning.

Conclusions:

• The proposed WSAD algorithm significantly reduces annotation effort for identifying anomalous CT slices.
• The WSAD algorithm demonstrates superior performance (higher AUC) compared to existing anomaly detection techniques.
• This approach offers a more efficient and effective solution for AI-assisted CT screening.