Self-supervised out-of-distribution detection-Metal implants and other anomaly

Area of Science:

• Artificial Intelligence in Medical Imaging
• Deep Learning for Anomaly Detection
• Medical Image Analysis

Background:

• Deep learning models struggle with external CT data due to artifacts from motion and implants.
• Supervised models are impractical for identifying diverse, unseen anomalies (out-of-distribution data).

Purpose of the Study:

• Develop an AI model to detect and identify anomalies/out-of-distribution data in abdominal-pelvis CT exams.
• Improve the performance of downstream AI applications using anomaly detection.

Main Methods:

• Proposed 2D and 3D generative architecture using Vector Quantized Variational Autoencoder (VQVAE) and Vision Transformer-Masked Autoencoder (VIT-MAE).
• Trained on over 2850 abdominal-pelvis CT volumes (adults >50 years) from Mayo Clinic.
• Tested on prospective and external datasets, including AbdominalCT-1k.

Main Results:

• Generative models achieved excellent results with negligible false positives, outperforming traditional methods.
• Prospective analysis showed an 86.11% true positive rate, handling under-documented anomalies.
• External validation on AbdominalCT-1k dataset yielded a 75.26% true positive rate.

Conclusions:

• The AI method effectively detects intra- and inter-class out-of-distribution data in abdominal CT images.
• The approach can assess CT dataset quality, providing actionable insights for data curation.
• The algorithm is valuable for secure healthcare collaborations and is available on GitHub.