Feasibility study of unsupervised anomaly detection using Wasserstein GAN in SPECT image.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Generative models

Background:

• Brain single-photon emission computed tomography (SPECT) is crucial for diagnosing neurological disorders.
• Accurate anomaly detection in SPECT images is essential for timely diagnosis.
• Current methods may be labor-intensive and prone to human error.

Purpose of the Study:

• To develop and assess the feasibility of an anomaly detection system for brain SPECT images.
• To utilize the Wasserstein generative model for enhanced image analysis.
• To investigate the system's potential to aid in clinical diagnosis.

Main Methods:

• Implemented a Wasserstein generative adversarial network (WGAN) based on optimal transport theory.
• Trained the WGAN using only healthy brain SPECT images for anomaly detection.
• Validated the system using a numerical phantom and clinical brain SPECT images, evaluating with receiver operating characteristic curves and area under the curve (AUC).

Main Results:

• The numerical phantom demonstrated a strong correlation between noise/signal levels and anomaly detection performance (AUC of 0.9994).
• A clear correlation between anomaly scores and detected anomalies was observed in clinical brain SPECT images.
• Subtraction images effectively highlighted abnormal regions.

Conclusions:

• The proposed Wasserstein generative model-based anomaly detection system is feasible for brain SPECT imaging.
• The system demonstrates high accuracy and potential for reducing the workload of human interpreters.
• This technology could improve the efficiency and reliability of SPECT image analysis.