Efficient Anomaly Detection with Generative Adversarial Network for Breast Ultrasound Imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Breast ultrasound is crucial for diagnosing breast masses.
• Accurate differentiation between normal tissue, benign, and malignant masses remains a challenge.
• Generative Adversarial Networks (GANs) offer potential for advanced image analysis.

Purpose of the Study:

• To evaluate the efficacy of GAN-based anomaly detection for diagnosing breast ultrasound images.
• To assess the model's ability to differentiate between normal tissue, benign masses, and malignant masses.
• To quantify the performance of the anomaly detection model using sensitivity, specificity, and AUC.

Main Methods:

• Retrospective collection of 531 normal breast ultrasound images.
• Data augmentation to create 6372 training images.
• Development of an efficient GAN-based anomaly detection model to calculate anomaly scores for lesion detection.

Main Results:

• Malignant masses exhibited significantly higher anomaly scores than benign masses (p < 0.001).
• Benign masses showed significantly higher anomaly scores than normal tissues (p < 0.001).
• The model achieved high sensitivity, specificity, and AUC values in distinguishing all tissue types, particularly for malignant masses.

Conclusions:

• GAN-based anomaly detection demonstrates high performance in detecting and diagnosing anomalous lesions in breast ultrasound.
• The model shows significant potential for improving the accuracy of breast mass diagnosis.
• This approach offers a promising tool for radiologists in breast cancer screening and diagnosis.