Individualized Statistical Modeling of Lesions in Fundus Images for Anomaly Detection

Area of Science:

• Ophthalmology and Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Anomaly detection in fundus images is challenging due to diverse lesion characteristics.
• Existing methods struggle with lesion variability and normal personalized variations (NPV).
• Reconstruction methods overlook lesion constraints, while separation methods fail to capture lesion individuality.

Purpose of the Study:

• To develop an improved anomaly detection method for fundus images.
• To address limitations of current reconstruction and separation-based approaches.
• To enhance the accurate identification of diverse lesions and reduce false positives.

Main Methods:

• Proposed a patch-based non-i.i.d. mixture of Gaussian (MoG) model for diverse lesion characterization.
• Introduced weighted Schatten p-norm for low-rank decomposition to refine background learning.
• Integrated individualized lesion modeling with background learning for improved distinction.

Main Results:

• The proposed method effectively models statistical variations and structural properties of diverse lesions.
• Weighted Schatten p-norm enhanced background accuracy and reduced false positives from NPV.
• Demonstrated superior performance compared to state-of-the-art methods on multiple databases.

Conclusions:

• The developed approach offers a robust solution for anomaly detection in fundus images.
• Accurate modeling of lesions and backgrounds improves diagnostic capabilities in retinal imaging.
• This method holds promise for advancing automated analysis of ocular pathologies.