Automated lesion detectors in retinal fundus images

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Vision

Background:

• Diabetic retinopathy (DR) is a leading cause of vision loss in diabetic patients.
• Early detection of DR is crucial for timely intervention and preventing blindness.
• Retinal fundus images are key for identifying early DR signs like microaneurysms, hemorrhages, and exudates.

Purpose of the Study:

• To develop and validate a novel automated system for detecting and diagnosing key retinal lesions associated with diabetic retinopathy.
• To improve the accuracy and efficiency of DR screening through advanced image processing techniques.

Main Methods:

• Proposed an automated system utilizing isotropic undecimated wavelet transform on the retinal image green channel.
• Derived novel contextual/numerical features based on Hessian multiscale analysis, variational segmentation, and cartoon+texture decomposition.
• Developed binary classifiers for microaneurysms, hemorrhages, and bright lesions, validated on 45,770 retinal fundus images.

Main Results:

• Individual lesion detectors achieved high performance: MA (93% sensitivity, 89% specificity), HEM (86% sensitivity, 90% specificity), BL (90% sensitivity, 97% specificity).
• The collective automated screening system demonstrated 95-100% sensitivity and 70% specificity on a per-patient basis.
• Comparative evaluation on public datasets indicated promising potential against existing techniques.

Conclusions:

• The developed automated system shows significant potential for accurate and efficient diabetic retinopathy screening.
• The novel feature extraction and classification methods contribute to improved detection of early DR signs.
• This technology can aid clinicians in preserving vision for diabetic patients through early diagnosis.