Retinal Lesion Detection With Deep Learning Using Image Patches.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Automated detection of retinal pathologies is crucial for early diagnosis and treatment.
• Limited training data poses a significant challenge for developing robust diagnostic models, especially for rare diseases.
• Current methods often rely on hard-coded feature extraction, limiting generalizability.

Purpose of the Study:

• To develop an automated method for localizing and discerning multiple types of retinal findings.
• To enable generalization to rare disease detection using limited training data.
• To avoid hard-coded feature extraction in the automated analysis of retinal images.

Main Methods:

• Two ophthalmologists verified 243 retinal images, creating 1324 labeled image patches.
• A standard convolutional neural network was trained to classify five classes: hemorrhages, microaneurysms, exudates, retinal neovascularization, and normal structures.
• A sliding window approach generated probability maps across entire retinal images.

Main Results:

• The method achieved a pixel-wise area under the receiver operating characteristic curve of 0.94 for microaneurysms and 0.95 for exudates.
• Lesion-wise area under the precision-recall curve was 0.86 for microaneurysms and 0.64 for exudates.
• Validation was performed on the eOphta dataset.

Conclusions:

• Regionally trained convolutional neural networks can effectively generate lesion-specific probability maps.
• These models can detect and distinguish subtle pathologic retinal lesions.
• The approach requires only a few hundred training examples per lesion, demonstrating its efficiency with limited data.