Abstract
Purpose :
Very recently, deep learning which is a new paradigm in machine learning, has shown significant generalisability across different image recognition tasks. This paper aims at applying deep learning method for automatic detection and classification of diabetic retinopathy (DR) features namely microaneurysms (MA), hard and soft exudates (ED), and hemorrhages (HM) in digital color fundus photographs.
Methods :
We used deep learning framework matconvnet [http://www.vlfeat.org/matconvnet/] to extract features of 4096 dimensionality from color fundus images. In accordance with many others we extracted deep features at the last convolutional layer and at the first fully connected layer. Then we used support vector machine (SVM) classifiers on top of these features. LibSVM [https://www.csie.ntu.edu.tw/~cjlin/libsvm/] toolbox was used with default parameters. Optic disk (OD) detection was also performed prior deep learning to avoid the ambiguous color feature match with exudates.
DIARETDB1 dataset (89 color fundus images (28 train images, 61 test images)) had been used for the experiment. For comparison, baseline method had been used in accordance with the evaluation protocol proposed in DIARETDB1.
Results :
Figure 1 shows the receiver operating characteristic (ROC) curves for the detection and classification of DR features. The weighted error rate (for R=1) of the proposed method (last conv. and first FC.) and baseline method are respectively (0.19, 0.17), 0.31 for microaneurysms. For hard and soft ED, and HM these values are respectively (0.22, 0.17), 0.16; (0.14, 0.21), 0.31; (0.20, 0.13), 0.50. Table 1 shows the effect of OD detection as a pre-processing step.
Conclusions :
A novel approach to MA, ED and HM detection using deep learning and SVM classification methods was proposed. The experimental results have shown that the proposed method significantly outperforms the baseline method, except for hard exudates. Overall OD detection improves the performance of the deep learning framework and provides significant improvement for soft-exudates detection.
This is an abstract that was submitted for the 2016 ARVO Annual Meeting, held in Seattle, Wash., May 1-5, 2016.