AI Medical Imaging: Revolutionizing Global Healthcare Access for Remote Diagnoses

In my previous article titled "Transforming Healthcare: A Glimpse into the AI Revolution", published as part of the newsletter, Innovate & Lead Chronicles, I shared my encounter with an AI Medical Scribe at a Doctor's office and discussed how it reduces the doctor’s administrative burden and significantly improves the patient experience.

Taking the discussion on AI revolutionizing healthcare further, this article discusses at a high level, my personal experience developing AI applications to diagnose diseases, and the tremendous benefit of serving the unserved and remote populations around the world.

Please note: If you are interested in reading a detailed in-depth technical whitepaper on this topic, where I discuss architecture, different AI-CNN models, data pre-processing, and results, please comment on this article and make a request below.

Journey into Radiology: Medical Images, Expert Diagnosis, and the Global Healthcare Disparities

A couple of years ago, a close relative suffering from back pain shared with me a copy of the MRI images and a doctor's report. I stared at the numerous MRI images for quite a while trying to first understand where the head or tail of the spine was, which side was front and back, and attempted to distinguish between muscle, bone, and nerves. Intriguingly, the radiologist had it all figured out and summarized it quite eloquently, "L5/S1 - disc extrusion indenting the traversing left S1 nerve root, producing left lateral recess and spinal canal stenosis. Left neural foraminal stenosis due to foraminal disc bulge and facet arthrosis.". Lacking expertise in this domain, I had no idea what all of that meant. I googled and watched YouTube videos, trying to transform myself into a radiologist in an hour, without any success.

You might have guessed what I trying to convey - you need an expert to interpret the results to make an accurate diagnosis. Various technologies such as X-rays, high-frequency sound waves used in ultrasonography, magnetic fields with radio waves in MRIs, and radiotracers for nuclear medicine, help the experts see inside our bodies and accurately diagnose tumors, cysts, internal injuries, arterial blockages, arthritis, pneumonia, spinal abnormalities, cancer, and other medical conditions.

The demand for highly trained and experienced doctors capable of accurate diagnoses is on the rise. Unfortunately, according to the World Health Organization, there is a global deficit of 4.3 million physicians, nurses, and health professionals, especially in developing countries. As an example, consider Diabetic Retinopathy, a complication of diabetes leading to blindness that poses a significant global health concern. With diabetes rates projected to rise from 463 million to 700 million by 2045, early detection becomes crucial to prevent blindness. Ophthalmologists are trained to examine fundus images (retinal images) and diagnose various eye ailments, including Diabetic Retinopathy. Unfortunately, these specialists are hard to find in developing countries, especially in remote areas.

Transforming Healthcare: My Journey in Developing an AI Medical Imaging Analysis Application for Diabetic Retinopathy Detection

Inspired by lectures at Stanford University by Ronjon Nag, I ventured into creating an AI-based medical image analysis application focused on detecting Diabetic Retinopathy (DR). I then built an application on a smartphone that can scan the fundus (retina) image with an external camera (as shown below) while the AI application detects the DR level in real time, as illustrated in the second picture below.

In the next version, this application can be possibly enhanced to connect to any available doctor on call, in the global network if the DR is detected. The remote doctor can cross-check and confirm the diagnosis, after which the application can schedule medical appointments at nearby healthcare facilities based on the physician's recommendation. This can further be seamlessly integrated with the Patient Health Record.

Critical Steps: Process Flow for Development

The complete process, as the flow chart depicts, is described below in seven steps:

1. Data Collection and Data Processing: The first vital step is to collect tens of thousands of high-quality retinal images each labeled to indicate the level of DR by a trained doctor. I discussed "Why is the data so important?" in another article titled "Navigating the AI Race...", and here also high-quality data is crucial for training the machine learning model to recognize patterns indicative of diabetic retinopathy and accurately diagnose the disease. The data must then be pre-processed (e.g. cropped to the same size with the same resolution) to ensure data consistency and accurate results.
2. Model Design and Training: The next step is to design an AI model and in this case a Convolutional Neural Network (CNN) model to effectively extract features from the retinal images. The model is trained using the labeled and curated dataset that I discussed above in the first step, and the parameters are fine-tuned to achieve the highest accuracy in detecting DR.
3. Model Deployment in the Cloud Platform: The pre-trained model is then deployed in the cloud for broader accessibility and scalability.
4. Condensing the Model for SmartPhone: Smartphones or most handheld devices have limited memory and storage. Hence the pre-trained model is condensed to fit and work on smartphones.
5. Deployment of the model into the mobile application: The model is then integrated with the developed application on the Smartphone.
6. Attach Fundus Camera: There are several smartphone-compatible cameras available in the market and the application is tested to ensure that it operates well with the camera.
7. Scan and get instant results: The application is now ready to take a picture of the retina and instantly process and detect the DR level of the patient. The results can be shared with the doctor for validation or further treatment.

With this successful implementation, let's explore the broader impact of AI applications in healthcare.

Critical Steps: Refining and Field Testing the Diabetic Retinopathy Model for Remote Deployment

Note that it is imperative to refine and field test the model before deploying it in the field to diagnose patients. Initial and ongoing feedback from experts is critical for continuous improvement and the success of the application.

Although AI-based diabetic retinopathy applications running a smartphone do not necessarily require highly trained doctors for usage, trained technicians are indispensable to capture high-quality retinal images and reduce the risk of generating false alarms potentially causing unnecessary panic among the patients.

Such AI applications for diagnosing diseases can be deployed in remote areas where access to qualified doctors is limited. With proper instructions, this screening can be as simple as checking the blood pressure or blood sugar. Once deployed, imagine how scalable and cost-effective this solution can be for underserved populations worldwide - enabling early detection and intervention to save the eyesight of millions of patients.

.

Conclusion & Next Steps

Beyond Diabetic Retinopathy, AI applications with trained CNN models have the potential to diagnose Neurological Disorders, Cancer cells, bacterial and viral infections, and other abnormalities. They promise to provide cost-effective solutions that can be deployed in developing countries.

As these technologies continue to advance, the collaboration between AI and medical professionals becomes a powerful tool in improving lives and advancing global health with AI leading the way.

Next Steps:

1. Share your perspective on using AI in diagnosing diseases in the comments below. Also, the next time you visit a doctor ask their perspective of using AI for Medical Diagnosis and treatment and share it here.
2. Extend the conversation by commenting on this article. Share your views or ask questions related to AI. Your input will be valuable to other readers of this newsletter and article.
3. Repost this article within your network and help disseminate this information.
4. Learn more: Comment below if you need additional information on AI-based medical diagnosis or more in-depth technical write-ups or code samples for such an application.