Fig 1.
Examples of various challenges associated with complex slit-lamp images.
(a) The different amplification factors of optical device. (b) The non-cooperation of patient and the different angle of the photographer. (c) The complications of pediatric cataract. (d) The noise of eyelashes and white highlights.
Fig 2.
Flow chart of the automatic diagnosis process for slit-lamp images.
The gray, blue and green block diagram represent the ROI of lens localization, classification and grading procedure respectively. (ROI: region of interest).
Fig 3.
Examples of slit-lamp images of three-degree grading.
From (a) to (c) are area, density, and location grading, in which the upper images represent mild disease and the lower images are severe. (d) The normal images for comparison.
Fig 4.
Examples of lens localization.
(a) Representative original images from Fig 1. (b) Edge detection using the Candy operator on the H component of the hue, saturation, and value (HSV) color space. (c) The regions of lens localization using the Hough transform correspond to (b). (d) Edge detection using the S component. (e) Final lens localization. (f) Cropped images for constructing slit-lamp datasets.
Fig 5.
The architecture of the deep convolutional neural network.
(a) Main layers and connections of the CNN that is employed in our study. It consists of five convolutional and overlapping max pooling layers, which are indicated by red rectangles and green rectangles respectively, followed by three fully-connected layers. (b), (d) Convolutional and overlapping max pooling operations are represented respectively. (c) The non-saturating activation function ReLU was represented. (ReLU: rectified linear units; CNN: convolutional neural network; SVM: support vector machine; ROI: region of interest; FC: full-connected operation; P1~P4: pixel value after pooling operation).
Fig 6.
The t-SNE maps of CNN and traditional features.
From (a) to (d) are the classification and three-degree grading: area, density, location, in which the upper images are maps of FC7 features in CNN network and the lower images are typical traditional features wavelet transformation. And the blue and red dots represent samples of two different categories. (t-SNE: t-distributed stochastic neighbor embedding; CNN: convolutional neural network; FC7: the seventh fully-connected layer).
Table 1.
Quantitative evaluation for pediatric cataract classification and grading results with different CAD models.
Fig 7.
The ROC curve and AUC value for pediatric cataract.
This figure contains the ROC curve and AUC value of (a) classification and three-degree grading (b) area, (c) density, and (d) location with different CAD models. (ROC: receiver operating characteristics curve; AUC: area under the curve; CAD: computer aided diagnosis; CNN: convolutional neural network; WAVE: wavelet transformation; LBP: local binary pattern; SIFT: scale-invariant feature transform; COTE: color and texture features).
Fig 8.
(a) The accuracy curve in respect to iterations. (b) The loss function value curve in respect to iterations.
Fig 9.
Heat map of the performance of RBF kernels with different values of c and gamma.
Different colors represent different levels of accuracy, and the accuracy increases from blue to red. (RBF: radial basis function; c: cost parameter of RBF; g: gamma parameter of RBF).
Table 2.
Comparison of SVM and softmax classifier with FC7 high-level features extracted by the CNN.