Abstract
Purpose :
To propose an automated decision support system for the ‘disease/no-disease’ grading of AMD relying of color normalization, deep feature extraction and support vector machine.
Methods :
A color normalization method was proposed to eliminate non-uniform illumination (step-1) and inter patient color variability (step-2). Step-1: the image was first transformed to HSV space where background subtraction method was applied on the V channel and then transformed back. Step-2: weighted Von Kries model [http://link.springer.com/chapter/10.1007/978-3-642-20404-3_13] was applied to perform color correction, where the mean (computed all over the images) R, G, B values of the optic disk and blood vessels, against the R, G, B values of the given image was used to compute the weights.
We used deep learning (DL) framework matconvnet [http://www.vlfeat.org/matconvnet/] on the color normalized images to extract deep features at the last convolutional (LC) layer and the first fully connected (FFC) layer. Support vector machine (SVM) [https://www.csie.ntu.edu.tw/~cjlin/libsvm/] classifiers were used on top of these features for the final grading.
Two publicly available AMD datasets namely ARIA, STARE were used for the experiment along with a private dataset of 4 AMD and 40 normal patients. All these datasets were merged into a single dataset. An experienced grader outlined the pathologies and graded the images. 50% of the images were used for training and rest for testing.
Results :
Figure 1 shows the outcome of color normalization, along with receiver operating characteristic (ROC) curves of the proposed system. Table 1 summarizes the overall findings in terms of sensitivity (SN), specificity (SP) and accuracy (Acc).
Conclusions :
A novel approach for the automated grading of AMD is proposed. The proposed system achieves an overall accuracy of 99.4% when deep features were learnt at the FFC layer and color normalization is performed prior to deep feature extraction. The proposed color normalization improves the overall accuracy of the DL framework by 16% and 3% when deep features are learnt at the LC and FFC layers respectively.
This is an abstract that was submitted for the 2017 ARVO Annual Meeting, held in Baltimore, MD, May 7-11, 2017.