Cooperative GAN: Automated tympanic membrane anomaly detection using a Cooperative Observation Network

Background and Objectives:

Recently, artificial intelligence (AI) has been applied to otolaryngology. However, existing supervised learning methods cannot easily predict data outside the learning domain. Moreover, collecting diverse medical data has become demanding owing to privacy concerns. Consequently, these limitations hinder the applications of AI in clinical settings. To address these issues, this study proposes a Cooperative Observation Network (CON), using an unsupervised anomaly detection approach. Anomaly detection is the process of identifying data patterns that deviate from the majority.

Methods:

For anomaly detection, the model is trained solely on normal data and calculates an abnormality score during the decoding process of the test via the reconstruction error. The calculated score is used to detect anomalies in the second step. Unlike traditional anomaly detection, the CON method does not rely on a decoding process. Instead, it detects anomalies in a single step using the discriminator of the Generative Adversarial Network. During the training process, the discriminator differentiates between the normal data distribution and artificially generated instances. However, these instances are obtained from a random distribution that does not overlap with the distribution of normal data. Consequently, the trained discriminator can recognize distributions outside the scope fo normal data. Additionally, we expand the diagnostic scope by utilizing two clinical variables: tympanic membrane endoscopic images and pure tone audiometry (PTA).

Results:

CON detects anomalies with a high accuracy of 96.75%. This includes cases with a normal tympanic membrane but with hearing loss, perforation, cholesteatoma, or retraction; cases with two co-existing diseases; and cases that require treatment but are difficult to diagnose with specific diseases. CON significantly reduces the computational load by approximately ten times compared with existing models while maintaining high accuracy and broadening diagnostic scope.

Conclusion:

This study successfully addresses the inherent limitations of supervised learning and anomaly detection, thereby enhancing the potential of AI-based disease detection in otolaryngology for practical clinical applications. The proposed methods can be seamlessly incorporated into medical machines for real-world clinical use owing to their low reliance on the computational load. Moreover, CON requires only a small amount of training data while maintaining the ability to diagnose a broad range of diseases with high accuracy. Therefore, it can effectively aid medical professionals in diagnosing in clinical scenarios, thereby increasing the efficiency of healthcare delivery.