A Deep Learning Method for Microaneurysm Detection in Fundus Images

Diabetic Retinopathy (DR) is the leading cause of blindness in the working-age population. Microaneurysms (MAs), due to leakage from retina blood vessels, are the early signs of DR. However, automated MA detection is complicated because of the small size of MA lesions and the low contrast between the lesion and its retinal background. Recently deep learning (DL) strategies have been used for automatic feature extraction and classification problems, especially for image analysis. In this paper, a Stacked Sparse Autoencoder (SSAE), an instance of a DL strategy, is presented for MA detection in fundus images. Small image patches are generated from the original fundus images. The SSAE learns high-level features from pixel intensities alone in order to identify distinguishing features of MA. The high-level features learned by SSAE are fed into a classifier to categorize each image patch as MA or non-MA. The public benchmark DIARETDB is utilized to provide the training/testing data and ground truth. Among the 89 images, totally 2182 image patches with MA lesions, serve as positive data, and another 6230 image patches without MA lesions are generated by a randomly sliding window operation, to serve as negative data. Without any blood vessel removal or complicated preprocessing operations, SSAE learned directly from the raw image patches, and automatically extracted the distinguishing features to classify the patches using Softmax Classifier. By employing the fine-tuning operation, an improved F-measure 91.3% and an average area under the ROC curve (AUC) 96.2% were achieved using 10-fold cross-validation.