Spatial-contextual variational autoencoder with attention correction for anomaly detection in retinal OCT images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Anomaly detection is crucial in various fields, including medical diagnostics.
• Existing reconstruction-based methods for anomaly detection have limitations due to single-constraint latent spaces.
• Retinal Optical Coherence Tomography (OCT) images require accurate anomaly detection for diagnosing eye conditions.

Purpose of the Study:

• To develop an advanced anomaly detection framework for retinal OCT images.
• To enhance the performance of anomaly detection by addressing limitations of previous methods.
• To accurately identify and localize abnormalities in retinal OCT scans.

Main Methods:

• Proposed a Spatial-Contextual Variational Autoencoder with Attention Correction (SCVAE-AC).
• Utilized a self-supervised segmentation network to isolate retinal regions, reducing background noise.
• Incorporated multi-dimensional and one-dimensional latent spaces to learn spatial and contextual features of normal images.
• Developed an ablation-based method for localizing anomalies and correcting anomaly scores.
• Constructed a novel anomaly score for improved separation of normal and abnormal images.

Main Results:

• The proposed SCVAE-AC framework demonstrated superior performance in anomaly detection on retinal OCT datasets.
• Effective elimination of background interference through self-supervised segmentation.
• Enhanced differentiation between normal and abnormal images due to improved latent space representation.
• Accurate localization of anomalous regions within the OCT scans.

Conclusions:

• The Spatial-Contextual Variational Autoencoder with Attention Correction is effective for anomaly detection in retinal OCT images.
• The method offers improved accuracy and localization capabilities compared to existing approaches.
• This technique holds promise for clinical applications in ophthalmology.