Method and system for choroid-scleral segmentation using deep learning with a choroid-scleral layer model

Patent metadata

Field	Value
Publication number	US-12575726-B2
Application number	US-202218085062-A
Country	US
Kind code	B2
Filing date	Dec 20, 2022
Priority date	Dec 20, 2021
Publication date	Mar 17, 2026
Grant date	Mar 17, 2026

Abstract

A System/Method/Device for segmenting the choroid-scleral layer from an optical coherent tomography (OCT) volume scan. The present system uses a deep learning machine model based on a neural network that include multiple convolution layers, but no deconvolution layers. Rather, the present neural network is based on a novel architecture based on the discrete cosine transform.

First claim

What is claimed:

1 . A method for segmenting a choroid-scleral layer from optical coherence tomography (OCT) image data of an eye, comprising:
collecting the OCT image data of the eye from an OCT system; submitting the OCT image data to a deep learning machine model based on a neural network, and producing an initial choroid-scleral layer segmentation, the neural network having:
a contracting path including a plurality of convolution layers followed by a max pooling layer; and
an expanding path following the contracting path, the expanding path including at least one discrete cosine transform (DCT) decoder; and
submitting the initial choroid-scleral layer segmentation to a classification module to identify and exclude, from the initial choroid-scleral layer segmentation, patch regions deemed to have incorrect segmentation and define a final choroid-scleral layer segmentation.

2 . The method of claim 1 , wherein the collected OCT image data is a volume scan, and the deep learning machine model uses a submission of a plurality of adjacent B-scans from the volume scan to produce the initial choroid-scleral layer segmentation.

3 . The method of claim 2 , wherein the plurality of adjacent B-scans from the volume scan are segmented simultaneously to define an 3D initial choroid-scleral layer segmentation with high correlation at a choroid region, and a deep learning model that enforces continuity of the segmented initial choroid-scleral layer between the plurality of adjacent B-scans.

4 . The method of claim 1 , wherein the collected OCT image data is a volume scan collected using an OCT system including:
a light source for generating a beam of light; a beam splitter having a beam-splitting surface for directing a first portion of the light into a reference arm and a second portion of the light into a sample arm; optics for directing the light in the sample arm to one or more locations on a sample; a detector for receiving light returning from the sample and reference arms and generating signals in response thereto; and a processor for converting the signals into OCT image data.

5 . The method of claim 1 ,
wherein the deep learning machine model is based on a neural network trained using a set of training input images and a set of training output images, and wherein each training output image is based on a corresponding training input image that has undergone attenuation correction, including removing of vessel shadows and improving contrast.

6 . The method of claim 1 , wherein the expanding path lacks any deconvolution module.

7 . The method of claim 1 , wherein the contracting path includes a relu-nonlinearity with 3×3 filter size followed by the max pooling layer.

8 . The method of claim 1 , wherein the contracting path defines an encoder, and wherein the max pooling layer determines a bottleneck size of the encoder that best reflects energy distribution of a discrete cosine transform decoder.

9 . The method of claim 8 , wherein the expanding path defines a decoder, and wherein a first step in a decoding operation of the decoder is an inverse cosine transform with the bottleneck size and followed by the plurality of convolutional layers.

10 . The method of claim 9 , wherein penultimate layers of the decoder includes zero-padding from the bottleneck size to the size of the originally submitted OCT image data combined with the discrete cosine transform to provide an image up-sampling function.

11 . The method of claim 1 , wherein the at least one discrete cosine transform (DCT) decoder computes the discrete cosine transform along the width and height dimensions.

12 . The method of claim 1 , wherein the OCT image data of an eye is a volume scan, wherein the volume scan is divided into 3D patches, and wherein the OCT image data submitted to the deep learning machine model is a 3D patch.

13 . The method of claim 1 , wherein the classification module uses choroid-scleral interface attributes, including an attribute of smooth surface with low frequency component, to identify incorrect segmentation patch regions of the initial choroid-scleral layer.

14 . The method of claim 1 , wherein the classification module identifies patch regions of incorrect segmentation based on a linear fit between a choroid-scleral interface and a smooth surface.

15 . The method of claim 14 , wherein the classification module identifies patch regions of incorrect segmentation based on one or more of outlier patches from a max curvature or arclength, unrealistic shapes defined as deviating beyond a predefined amount from a library of acceptable shapes, and choroid-scleral interface modeling based on polynomial, Zernike, or Gridfit.

16 . D) A system for segmenting a choroid-scleral layer from an optical coherence tomography (OCT) volume scan image data of an eye, comprising:
a light source for generating a beam of light; optics for directing the light to one or more locations on a sample including the eye; a detector for receiving light returning from the sample and generating signals in response thereto; and a processor operably coupled to the detector and configured to: define the volume scan image data from the generated signals, submit the volume scan image data to a deep learning machine model based on a neural network, and producing an initial choroid-scleral layer segmentation, wherein the neural network comprises: a contracting path including a plurality of convolution layers followed by a max pooling layer; and an expanding path following the contracting path, the expanding path including at least one discrete cosine transform (DCT) decoder; and submitting the initial choroid-scleral layer segmentation to a classification module to identify and exclude, from the initial choroid-scleral layer segmentation, patch regions deemed to have incorrect segmentation and define a final choroid-scleral layer segmentation.

17 . The system of claim 16 , wherein a plurality of adjacent B-scans from the volume scan image data are segmented simultaneously to define an 3D initial choroid-scleral layer segmentation with high correlation at a choroid region, and a deep learning model that enforces continuity of the segmented initial choroid-scleral layer between the plurality of adjacent B-scans.

18 . The system of claim 16 , wherein the expanding path lacks any deconvolution module.

19 . The system of claim 16 , wherein the contracting path defines an encoder, and the max pooling layer determines a bottleneck size of the encoder that best reflects the energy distribution of the discrete cosine transform decoder.

20 . The system of claim 19 , wherein the expanding path defines a decoder, and a first step in a decoding operation of the decoder is an inverse cosine transform with the bottleneck size and followed by a plurality of convolutional layers.

Assignees

• Zeiss Carl Meditec Inc

Inventors

• Bagherinia Homayoun
• Omlor Lars

Classifications

• G06T7/0012

  Biomedical image inspection · CPC title

• G06T2207/20081

  Training; Learning · CPC title

• G06T2207/20084

  Artificial neural networks [ANN] · CPC title

• G06T2207/30041

  Eye; Retina; Ophthalmic · CPC title

• G06T2207/10101

  Optical tomography; Optical coherence tomography [OCT] · CPC title