A recent Scientific Reports study developed DiaNet v2, an updated form of previously developed DiaNet, which was the first deep learning-based model to diagnose diabetes using retinal images.
Study: DiaNet v2 deep learning based method for diabetes diagnosis using retinal images. Image Credit: LALAKA/Shutterstock.com
Background
Diabetes mellitus (DM) is a metabolic disorder that is associated with long-term morbidity and mortality.
There are two types of DM, namely, type 1 DM (DM-1) and type 2 DM (DM-2). In comparison to DM-1, DM-2 is more commonly prevailing worldwide.
Millions of people worldwide are affected by DM, which is expected to reach 136 million by 2045. Early detection of this metabolic condition significantly impacts treatment and prevention.
Several tests, such as random plasma glucose (RPG), fasting plasma glucose (FPG), oral glucose tolerance tests (OGTT), and hemoglobin A1c (HbA1c), are performed to detect DM.
It must be noted that several limitations have been identified for each of the aforementioned tests. For instance, FPG tests have lower sensitivity, and a World Health Organization report stated that FPG has missed around 30% of diabetes diagnoses.
HbA1c results are also affected by different types of anemia or hemoglobinopathy, which could impact the diagnosis.
Considering the limitations in the available diabetes detection methods and their high prevalence rate, developing an alternative, cost-effective method with higher accuracy and sensitivity is important.
Previous studies have detected several alternative ways of diabetes detection that include the use of retinal images, electrocardiography (ECG), and breath tests.
As mentioned before, DiaNet had been previously developed as an alternative method to detect DM.
This deep learning-based model detected the metabolic disorder using retinal images and exhibited 84% accuracy in distinguishing diabetic individuals from non-diabetics.
About the study
This study used large cohorts from the Qatar Biobank (QBB) and Hamad Medical Corporation (HMC), the largest healthcare provider in Qatar, to improve DiaNet’s prediction capacity for diabetes.
The DiaNet v2 was developed using more than 5,000 retinal images. It must be noted that the proposed VGG-11-based DiaNet v2 model exhibited greater performance than DenseNet-121, ResNet-50, EfficientNet, and MobileNet_v2. VGG-11 network was trained with ImageNet, comprising an output of 1,000 neurons in its final layer.
A workstation comprising 12th Gen Intel(R) Core (TM) i7-127,00KF, with 128 GB RAM and GeForce RTX 3090 GPU, was used for all experiments.
In comparison to DiaNet v1, DiaNet v2 was trained using the combined dataset from QBB and HMC.
Study findings
A total of 15,011 images were obtained, among which 7,515 images were of diabetics and 7,496 were of non-diabetics or healthy controls.
The new model achieved over 92% accuracy in differentiating people with diabetes from the healthy control group, which is a significant achievement compared to the previous model.
The performance of DiaNet v2 was validated using the large-scale HMC and QBB dataset, which further confirmed retinal images are an excellent source to detect diabetes.
Retinal images from the QBB dataset lacked information about pre-existing and ocular pathologies. To overcome this data-related shortcoming, HMC data was integrated as it contained relevant information documented by ophthalmologists.
Retinal images of people with diabetes exhibited a range of pathologies, such as vitreous hemorrhage and microaneurysm, which is a consequence of being diabetic.
A diabetic eye also develops mild non-proliferative diabetic retinopathy (NPDR), an earlier stage of diabetic retinopathy (DR). The study cohort also comprised images of non-diabetic eyes with glaucoma.
These images were used to train the gender-stratified version of DiaNet v2. Interestingly, a higher accuracy in diabetic detection was observed in female participants.
Future studies must address this gender disparity to obtain a superior model for diabetes detection, irrespective of gender differences.
The age-stratified analysis revealed superior accuracy of VGG-11 across all age groups; however, the highest accuracy was achieved in age groups between 18 and 39 years, followed by 40 and 59 years.
The performance of the DiaNet v2 model was hindered in the 60–90 age group due to the smaller control group size. This finding reflects the importance of a balanced dataset for accurate prediction.
A Class Activation Map (CAM) analysis indicated the regions within the retinal image that influence the predictions of the DiaNet v2 model. These regions are associated with macula, optic disc, and areas linked to DR development.
The CAM analysis provided evidence of systematic conditions, such as ischemic heart disease, hypertension, and diabetes.
Conclusions
The current study revealed the potential of deep learning models based on retinal images in the diagnosis of diabetes.
The diaNet v2 model could be used as an effective, alternative, reliable, and non-invasive tool to diagnose diabetes. In the future, multi-modal approaches must be implemented to improve the model performance, which must be validated before being used in the real world.
Food additive emulsifiers linked to increased risk of type 2 diabetes