Anomaly guided segmentation: Introducing semantic context for lesion segmentation in retinal OCT using weak context...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Automated lesion detection in retinal optical coherence tomography (OCT) scans is crucial for clinical applications like diagnosis and treatment monitoring.
• Current segmentation models face challenges with complex lesions, variable image quality, and diverse disease presentations in real-world OCT data.
• Existing improvement techniques have not explored incorporating semantic context via anomaly detection.

Purpose of the Study:

• To investigate the effectiveness of integrating anomaly detection models with standard segmentation models for improved retinal lesion detection.
• To demonstrate that incorporating weak anomaly labels can enhance segmentation performance without additional manual annotation.
• To validate a generic anomaly-guided segmentation approach across multiple retinal OCT datasets and lesion types.

Main Methods:

• Developed a strategy to use a separate anomaly detection model to identify potential lesions.
• Integrated the output masks from the anomaly detection model as an additional class during the training of standard segmentation models.
• Trained and evaluated the combined model on two in-house and two publicly available retinal OCT datasets.

Main Results:

• Consistently improved lesion segmentation results when incorporating weak anomaly labels into standard segmentation models.
• Demonstrated enhanced performance across multiple lesion targets and diverse datasets.
• Validated the generic applicability of the anomaly-guided segmentation approach.

Conclusions:

• Anomaly-guided segmentation is a promising strategy to enhance automated lesion detection in retinal OCT scans.
• This method provides valuable semantic context, improving segmentation accuracy without the need for extensive manual labeling.
• The approach offers a practical tool for improving the robustness and performance of current lesion detection models.