Granted Patent B2

US 12,361,546 · App. 17/966,607 · Granted Jul 15, 2025

Inventors: Hyoung Uk Kim (Anyang-si, KR); Gu Yong Kim (Anyang-si, KR)

Assignee: HUVITZ CO., LTD.

G06T7/0012A61B3/0025A61B3/1005A61B3/102A61B3/1225G06T7/13G06T7/33G06T2207/10101G06T2207/20081G06T2207/20084G06T2207/30041

Abstract

Claims (18)

1. A method of measuring a retinal layer comprising a retinal nerve fiber layer, the method comprising:

step S 10 of obtaining an OCT image of a retina;

step S 12 of detecting a reference boundary line indicating a retinal layer in the obtained OCT image;

step S 14 of obtaining an aligned OCT image by aligning a vertical position of each column of the OCT image so that the detected reference boundary line becomes a baseline;

step S 20 of predicting retinal layer regions from the aligned OCT image;

step S 22 of calculating boundary lines between the predicted retinal layer regions; and

step S 30 of restoring the calculated boundary lines to positions of the boundary lines of the retinal layer of the original OCT image by aligning the vertical positions of the calculated boundary lines of the retinal layer for each column so that the baseline becomes the reference boundary line again; and

step S 32 of measuring a thickness of the retinal nerve fiber layer from the boundary line positions between the respective restored retinal layers,

wherein the reference boundary line is a boundary line between a vitreous body above the retina and an inner surface of the retina, and the baseline is a reference line that flattens the reference boundary line so as to be able to reduce measurement errors caused by irregular bending of the retinal layer, and

wherein the prediction of the retinal layer regions and the calculation of the boundary lines between the retinal layer regions are performed by a deep neural network trained by using (i) an aligned OCT image in which the reference boundary line is changed to be aligned to the baseline and (ii) a boundary line data of retinal layers obtained by analyzing the aligned OCT image, wherein training of the deep neural network is performed by:

step S 50 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 52 of detecting a reference boundary line indicating a retinal layer in the OCT image and label image inputted;

step S 54 of aligning a vertical position of each column of the OCT image and the label image so that the detected reference boundary line becomes a baseline;

step S 62 of predicting a position of the retinal layer from the OCT image when the OCT image and the label image are aligned;

step S 64 of calculating a prediction error by comparing the predicted position of the retinal layer and the label image; and

step S 66 of updating a weight to be predicted for each layer according to the calculated prediction error.

2. The method of measuring a retinal layer of claim 1 , further comprising:

a step of overlaying the restored boundary lines of the retinal layer on the original OCT image and displaying them to a user.

Assignments (1)

Priority Claims (1)

Continuity (1)

References Cited (11)