Method for differentiating retinal layers in oct image
A method for differentiating retinal layers includes obtaining an optical coherence tomography retinal layer image of an eye to be examined and inputting it into a deep neural network, calculating probability P that each pixel is included in each layer from the inputted layer image, calculating probability P1(x,y) that a pixel is in or above (k−1)-th layer and probability P2(x,y) that the pixel is in or below k-th layer from probabilities P(x,y) that the pixel is included in each layer, calculating probability F k−1,k (x,y) that the pixel becomes a boundary line between the (k−1)-th layer and the k-th layer from P1( x,y ) and P2( x,y ), obtaining a node cost C k−1,k (x,y) of the pixel by multiplying F k−1,k (x,y) with an image brightness change ΔI at the pixel, and determining the boundary line between the (k−1)-th and k-th layers by connecting coordinates (x,y) of the pixels where C k−1,k (x,y) calculated for each position x is the maximum.
1. A method of determining boundary lines between retinal layers, comprising:
step S 20 of obtaining an optical coherence tomography (OCT) retinal layer image of an eye to be examined and inputting it into a deep neural network;
step S 22 of calculating a probability P that each pixel is included in each layer from the inputted layer image;
step S 24 of calculating a probability P 1 ( x, y ) that a pixel at the coordinate (x, y) is in or above (k−1)-th layer and a probability P 2 ( x, y ) that the pixel is in or below k-th layer from probabilities P (x, y) that the pixel at the coordinate (x, y) is included in each layer, wherein the probability P 1 ( x, y ) and the probability P 2 ( x, y ) are calculated by steps of:
calculating a likelihood L 1 that the pixel at position (x, y) is located above a boundary line between the (k−1)-th layer and the k-th layer, the likelihood L 1 being determined by summing probabilities, output by the deep neural network, that the pixel is classified as belonging to one of retinal layers C uppers above the boundary line, and a likelihood L 2 that the pixel is located below the boundary line, the likelihood L 2 being determined by summing probabilities, output by the deep neural network, that the pixel is classified as belonging to one of retinal layers C lowers below the boundary line, using Equation 1 below:
L 1 ( x,y )Σ i∈cuppers Pl ( x,y ), L 2( x,y )=Σ i∈clowers pl ( x,y ) Equation 1:
obtaining the probability P 1 and the probability P 2 from the likelihood L 1 and the likelihood L 2 using Equation 2 below:
P
1
(
x
,
y
)
=
L
1
(
x
,
y
)
L
1
(
x
,
y
)
+
L
2
(
x
,
y
)
,
Equation
2
P
2
(
x
,
y
)
=
L
2
(
x
,
y
)
L
1
(
x
,
y
)
+
L
2
(
x
,
y
)
.
step S 26 of calculating a probability F k−1,k (x, y) that the pixel becomes a boundary line between the (k−1)-th layer and the k-th layer from the probability P 1 ( x, y ) that the pixel is in or above the (k−1)-th layer and the probability P 2 ( x, y ) that the pixel is in or below the k-th layer;
step S 28 of obtaining a node cost C k−1,k (x, y) of the pixel by multiplying the probability F k−1,k (x, y) that the pixel at the position (x, y) becomes the boundary line between the (k−1)-th layer and the k-th layer with an image brightness change ΔI at the pixel; and
step S 30 of determining the boundary line between the (k−1)-th layer and the k-th layer by connecting the coordinates (x, y) of the pixels where the node cost C k−1,k (x, y) calculated for each position x is the maximum.
2. The method of claim 1 , wherein the deep neural network is trained by:
step S 10 of inputting a training data set of (i) an OCT retinal cross-sectional image and (ii) a label image created for the OCT retinal cross-sectional image;
step S 12 of predicting positions of the retinal layers from the OCT image;
step S 14 of computing a prediction error (loss) by comparing the predicted positions of the retinal layers and the label image; and
step S 16 of updating a weight to be predicted for each layer according to the computed prediction error.
3. The method of claim 1 , wherein the image brightness change ΔI at the pixel is computed by Equation 5 below, in which I is a brightness value at the corresponding coordinates, and Δy represents a distance to an adjacent pixel:
Δ I ( X,Y )= I ( x,y+Δy )− I ( x,y ) Equation 5:
4. The method of claim 1 , further comprising:
a step S 32 of determining all boundary lines that differentiate respective retinal layers by repeating steps S 24 to S 30 for all the boundary lines that differentiate the respective retinal layers in the OCT image.