Change detection based on unsupervised sparse representation for fundus image pair.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Longitudinal fundus image comparison is crucial for detecting ocular changes.
• Pixel-by-pixel comparison methods are susceptible to illumination variations, obscuring true changes.
• Robust change detection is needed to overcome these limitations in ophthalmology.

Purpose of the Study:

• To propose a novel unsupervised change detection method for fundus image pairs.
• To address and correct for illumination variations in longitudinal fundus images.
• To enhance the accuracy and robustness of change region detection in ophthalmology.

Main Methods:

• Developed an unsupervised change detection method utilizing sparse representation classification (SRC).
• Constructed a local background dictionary from reference image patches.
• Reconstructed current image patches using sparse representation for background subtraction and change detection.

Main Results:

• The SRC method demonstrated high performance with AUC of 0.9858 and mAP of 0.8647 for small lesions.
• A fusion method combining IRHSF and SRC achieved superior results (AUC: 0.9892, mAP: 0.9692) for larger change regions.
• The proposed SRC method proved more robust to illumination variations than RPCA and more effective than pixel-wised differencing.

Conclusions:

• The SRC-based method effectively detects changes in fundus images by automatically correcting for illumination variations.
• This approach offers improved robustness and accuracy compared to existing methods for longitudinal ocular image analysis.
• The findings support the utility of SRC for reliable change detection in ophthalmological applications.