A Novel Approach to Cut Costs Associated With Diabetic Retinopathy Diagnoses Utilizing Deep Learning

Rishi Bathala

Background

According to the IDF Diabetes Atlas, around 537 million adults live with diabetes. This number is only going to continue to get worse, as by 2045, 783 million people will have diabetes. With this epidemic on the rise, a specific manifestation of diabetes in the type of a retinal degeneration known as diabetic retinopathy affects people’s eyes. However, many cases of diabetic retinopathy go undiagnosed due to the high cost of medical care affecting those of low-income. While, in the status quo, practitioners have existing AI-based approaches, a key problem with them is the lack of interpretability which makes it hard for doctors to understand why an AI system makes a particular diagnosis. In order to construct an accurate model capable of diagnosis, I developed a highly accurate and interpretable diagnostic model of diabetic retinopathy. I achieved this by training a type of convolutional neural network known as ResNet-50 on a dataset which contained 3,662 fundus photographs of the retina. This dataset included 5 stages of diabetic reitnopathy in patients. To address the issue of interpretability, I utilized a technique known as Grad-CAM, which highlights the regions of the image that contributed the most to the model's decision of categorizing an image, helping to see through the model’s eyes. I found that the ResNet-50 model was the most accurate at predicting diabetic retinopathy diagnosis, with a final accuracy of 75%. The Grad-CAM method yielded explainable feature visualizations which will be useful in showing a medical provider the affected regions of the retina which are leading to a particular diagnosis. This work demonstrates the effectiveness of deep learning-based diagnostic tools and has important implications for future practices.

How can this approach help people of low income?

Accessibility: By utilizing AI and computer vision technology, my research is a step in the right direction towards a cost-effective solution for screening and diagnosing diabetic retinopathy to those who especially need it.
Early detection: The ability of my convolutional neural network accurately detecting different stages of diabetic retinopathy can enable early diagnosis, even in resource-limited settings. Early detection is crucial for preventing vision loss.
Interpretability: The use of Grad-CAM to provide explanations and visualizations of the affected regions in the retinal images can help medical practitioners with an extra helping hand, taking burden off of them allowing for more precise and faster diagnosis.
Cost savings: By automating the screening and diagnostic process, my approach has the potential to reduce the overall cost of diabetic retinopathy management, making it more accessible and affordable for low-income patients.

Results

This is a technique known as Grad-CAM (Gradient-weighted Class Activation Mapping) being applied to a photograph where diabetic retinopathy isn’t present. We see that the model classified the image as no diabetic retinopathy. If we take a look at the image to the far right the model highlighted the most influential region in a bright red. We can verify that this patient has no diabetic retinopathy as there are no yellow splotches visible.

What is happening in the image to the right?

The leftmost image is the unaltered original image. In the middle, the heatmap of the image showcases impactful regions in shades of bright yellow, green and turquoise. On the far right is the heatmap overlaid onto the original image. Regions highlighted in red, orange, yellow and green correspond to regions that strongly contribute to the model’s decision. On the other hand, regions highlighted in blue and purple don’t influence the model’s decision by a lot. In this image we see that he model most likely classified this image

This is a technique known as Grad-CAM (Gradient-weighted Class Activation Mapping) being applied to a photograph where moderate diabetic retinopathy is present. The model accurately classified this image as one with moderate diabetic retinopathy. To verify this, if we take a closer look at the center of the fundus scan image, we again see very distant yellow splotches, hence, why the model classified this image as one with moderate diabetic retinopathy.

What is happening in the image to the right?

The leftmost image is the unaltered original image. In the middle, the heatmap of the image showcases impactful regions in shades of bright yellow, green and turquoise. On the far right is the heatmap overlaid onto the original image. Regions highlighted in red, orange, yellow and green correspond to regions that strongly contribute to the model’s decision. On the other hand, regions highlighted in blue and purple don’t influence the model’s decision by a lot. In this image we see that he model most likely classified this image

This is a technique known as Grad-CAM (Gradient-weighted Class Activation Mapping) being applied to a photograph where diabetic retinopathy is present. The model accurately classified the image as one with proliferate diabetic retinopathy. If we take a closer look at the center of the fundus scan image, we see a lot of yellow splotches coating the entire image, hence, why the model classified this image as one with proliferate diabetic retinopathy.

What is happening in the image to the right?

The leftmost image is the unaltered original image. In the middle, the heatmap of the image showcases impactful regions in shades of bright yellow, green and turquoise. On the far right is the heatmap overlaid onto the original image. Regions highlighted in red, orange, yellow and green correspond to regions that strongly contribute to the model’s decision. On the other hand, regions highlighted in blue and purple don’t influence the model’s decision by a lot. In this image we see that he model most likely classified this image.

This is a technique known as Grad-CAM (Gradient-weighted Class Activation Mapping) being applied to a photograph where mild diabetic retinopathy present. We see that the model classified the image as one with mild diabetic retinopathy. As we take a closer look at the image to the left, we see very distant yellow splotches visible. This happens because of the excess lipoproteins leaking from damaged retinal vessels. The model highlights this area in red validating why it classified this image as mild diabetic retinopathy.

What is happening in the image to the left?

This is a technique known as Grad-CAM (Gradient-weighted Class Activation Mapping) being applied to a photograph where severe diabetic retinopathy is present. The model accurately classified the image as one with severe diabetic retinopathy. If we take a closer look at the center of the fundus scan image, we see a lot of yellow splotches much more visible than the the image with moderate diabetic retinopathy, hence, why the model classified this image as one with severe diabetic retinopathy.

What is happening in the image to the left?