US-12620087-B2 - Dianet: a deep learning based architecture to diagnose diabetes using retinal images only

Abstract

A method of training a convolutional neural network model to predict diabetes from an image of a retina is provided. The method of training a convolutional neural network includes processing a first dataset, wherein processing the first dataset comprises: extracting a circular region from a retinal image, resizing the circular region, cropping the circular region, and placing the circular region onto a black background; training an initial model using a second dataset to yield a first model; training the first model using a third dataset to yield a second model; and training the second model using the first dataset to yield a third model.

Inventors

• Tanvir Alam

Assignees

• QATAR FOUNDATION FOR EDUCATION, SCIENCE AND COMMUNITY DEVELOPMENT

Dates

Publication Date

20260505

Application Date

20230109

Claims (15)

1. 1 . A method of training a convolutional neural network model to predict diabetes from an image of a retina, comprising: processing a first dataset, wherein processing the first dataset comprises: extracting a circular region from a retinal image; resizing the circular region; cropping the circular region; and placing the circular region onto a black background; training an initial model using a second dataset to yield a first model; training the first model using a third dataset to yield a second model; and training the second model using the first dataset to yield a third model-, wherein the third dataset comprises a plurality of retinal images labeled based on the severity of diabetic retinopathy.
2. 2 . The method of training a convolutional neural network model to predict diabetes from the image of a retina of claim 1 , wherein processing the first dataset outputs an image having a retina with a radius of 300 pixels.
3. 3 . The method of training a convolutional neural network model to predict diabetes from the image of a retina of claim 1 , wherein the initial model is a DenseNet model.
4. 4 . The method of training a convolutional neural network model to predict diabetes from the image of a retina of claim 3 , wherein the DenseNet model is a 121-layer variant DenseNet model.
5. 5 . The method of training a convolutional neural network model to predict diabetes from the image of a retina of claim 1 , wherein the first dataset comprises a plurality of retinal images including a first group of retinal images and a second group of retinal images.
6. 6 . The method of training a convolutional neural network model to predict diabetes from the image of a retina of claim 5 , wherein the first group of retinal images relates to a control group and the second group of retinal images relates to a diabetes group.
7. 7 . A convolutional neural network model architecture to predict diabetes from an image of a retina, comprising: a DenseNet-121 backbone; a final layer outputting a predictive label corresponding to a diabetic prediction or a non-diabetic prediction; a pair of pooling layers; and a first composite layer, a second composite layer, and a third composite layer, wherein the pair of pooling layers comprise a global average pooling layer and a global max pooling layer.
8. 8 . The convolutional neural network model architecture of claim 7 , wherein the final layer is a single neuron layer.
9. 9 . The convolutional neural network model architecture of claim 7 , wherein the first composite layer comprises a first sequence of a plurality of layers, the second composite layer comprises a second sequence of a plurality of layers, and the third composite layer comprises a third sequence of a plurality of layers.
10. 10 . The convolutional neural network model architecture of claim 9 , wherein the first sequence of the plurality of layers comprises a batch normalization layer, a dropout layer, a linear layer, and a rectified linear unit layer.
11. 11 . The convolutional neural network model architecture of claim 9 , wherein the second sequence of the plurality of layers comprises a batch normalization layer, a dropout layer, a linear layer, and a rectified linear unit layer.
12. 12 . The convolutional neural network model architecture of claim 9 , wherein the third sequence of the plurality of layers comprises a batch normalization layer, a dropout layer, and the final layer for outputting the predictive label.
13. 13 . A method of training a convolutional neural network model to predict diabetes from the image of a retina, comprising: processing a first dataset and a second dataset, wherein processing the first dataset and the second dataset comprises: extracting a circular region from a retinal image; resizing the circular region; cropping the circular region; and placing the circular region onto a black background; training a DenseNet-121 model using an third dataset to yield a first model; training the first model using the second dataset to yield a second model; and training the second model using the first dataset to yield a third model, wherein the third model has an architecture comprising: a DenseNet-121 backbone; a global average pooling layer; a global max pooling layer; a first composite layer; a second composite layer; and a third composite layer.
14. 14 . The method of training a convolutional neural network model to predict diabetes from the image of a retina of claim 13 , wherein the third composite layer comprises: a batch normalization layer, a dropout layer; and a final layer outputting a predictive label corresponding to a diabetic prediction or a non-diabetic prediction.
15. 15 . The method of training a convolutional neural network model to predict diabetes from the image of a retina of claim 13 , wherein the first composite layer comprises: a batch normalization layer; a dropout layer; a linear layer; and a rectified linear unit layer.

Description

CROSS-REFERENCES TO RELATED APPLICATIONS The present disclosure claims priority to U.S. Provisional Patent Application 63/298,473 titled “DIANET: A DEEP LEARNING BASED ARCHITECTURE TO DIAGNOSE DIABETES USING RETINAL IMAGES ONLY” having a filing date of Jan. 11, 2022, the entirety of which is incorporated herein. BACKGROUND Diabetes is one of the leading fatal diseases globally, putting a huge burden on the global healthcare system. Diabetes mellitus or diabetes is considered a collection of metabolic conditions that can predominantly be described by hyperglycemia rising from the deficiency in insulin discharge. The prolonged hyperglycemia of diabetes is correlated with long-term impairment and collapse of heart, kidneys, and microvascular circulation of the retina. Among diabetic individuals in the USA, almost 30% of them have the tendency of growing diabetic retinopathy, a common complication for diabetic patients, which may lead to blindness. Diabetes may adversely affect the vascular system of the retina causing structural change of it. As the changes in vascular structure in retina can provide visual cues for diabetes, most clinical guidelines recommend annual retinal screening for diabetic patients through retinal fundus images or dilated eye examinations. Alternatively, these retinal images can be used to detect diabetes, but it requires subjective judgement from the ophthalmologist, and it might be time consuming as well. Automatic retinal image-based diabetes diagnosis in a clinical setting could alleviate the workload of the ophthalmologist as well as screen a large number of patients objectively within a short amount of time. Current research has been conducted, which include (1) detecting diabetic retinopathy from retinal images and (2) diagnosing of diabetes based on clinical markers e.g., HbA1c, Glucose. However, current research has not addressed the task of detecting diabetes using retinal images from a holistic point of view, independent of the presence of diabetic retinopathy. Thus, improved non-invasive diabetes screening solutions are needed. SUMMARY The present disclosure generally relates to a method of training a convolutional neural network model to predict diabetes from the image of a retina. In light of the present disclosure, and without limiting the scope of the disclosure in any way, in an aspect of the present disclosure, which may be combined with any other aspect listed herein unless specified otherwise, a method of training a convolutional neural network model to predict diabetes from the image of a retina is provided. In an aspect of the present disclosure, which may be combined with any other aspect listed herein unless specified otherwise, the method of training a convolutional neural network model to predict diabetes from the image of a retina, comprising: processing a first dataset, wherein processing the first dataset comprises: extracting a circular region from a retinal image; resizing the circular region; cropping the circular region; and placing the circular region onto a black background; training an initial model using a second dataset to yield a first model; training the first model using a third dataset to yield a second model; and training the second model using the first dataset to yield a third model. In an aspect of the present disclosure, which may be combined with any other aspect listed herein unless specified otherwise, processing the first dataset outputs an image having a retina with a radius of 300 pixels. In an aspect of the present disclosure, which may be combined with any other aspect listed herein unless specified otherwise, the initial model is a DenseNet model. In an aspect of the present disclosure, which may be combined with any other aspect listed herein unless specified otherwise, the DenseNet model is a 121-layer variant DenseNet model. In an aspect of the present disclosure, which may be combined with any other aspect listed herein unless specified otherwise, the first dataset comprises a plurality of retinal images including a first group of retinal images and a second group of retinal images. In an aspect of the present disclosure, which may be combined with any other aspect listed herein unless specified otherwise, the first group of retinal images relates to a control group and the second group of retinal images relates to a diabetes group. In an aspect of the present disclosure, which may be combined with any other aspect listed herein unless specified otherwise, the third dataset comprises a plurality of retinal images labeled based on the severity of diabetic retinopathy. In an aspect of the present disclosure, which may be combined with any other aspect listed herein unless specified otherwise, the convolutional neural network model architecture to predict diabetes from an image of a retina, comprises: a DenseNet-121 backbone; and a final layer outputting a predictive label corresponding to a diabetic prediction or a non-diabetic