Today I dive into the world of an engineer bridging the gap between Artificial intelligence and Healthcare. I met Ajay Shekar, a Master’s student in Computer Science, on a group trip to Nashville and the Great Smoky Mountains, where I discovered his quirky personality and down-to-earth nature. I had no idea about the potential impact of his work as a scientist, and today I have the honor of sharing his work.
For the past year, Shekar has been working on applying machine learning and deep learning techniques on medical image datasets. He wants to create a tool that can help predict Parkinson’s Disease (PD) with the use of MRI and SPECT images, and he’s currently working on a paper titled “Residual CNNs for Predicting Parkinson’s Disease from SPECT Images.” He shares that one of the interesting bits is the kind of data he works with. Each patient SPECT image is approximately of the dimensions 90x90x180. You can think of it as a images stacked on top of each other. The stacked images give you a volume, which is what patients see as their 3D MRI scan. What is interesting with this kind of data is how they extract features in order to build a classifier (how discriminative your feature set is will decide the performance of your classifier). They employ the use of 3D Convolutional Neural Networks in order to extract a feature set that spreads across a volume of an image. They use this feature set for classification. Currently, just using a single modality of SPECT images, they are able to achieve ~90% accuracy in the classification of PD. The next step would be to incorporate features derived from different modalities of images (such as MRIs) in the hopes of achieving a higher accuracy.
Another aspect of Shekar’s work is to see what the trained neural network has learnt (still an open problem) and see if it actually correlates to what the general medical community uses towards the classification of PD. He found that the portion of the SPECT image that contributes highest towards the classification of PD is the same as what other researchers have established. The current state of the art methods achieve ~97% accuracy. There is still ways to go.
Shekar’s interest in research stems from his desire to solve challenging and impactful problems. He feels that getting involved in research seemed like a logical progression given his background knowledge. Looking back at his education, he realized that he had worked on other innovative and open problems that did not have a direct solution. These kind of problems instigated a need to look at what the community in general has established as “best methods” and work towards solutions that create a better baseline, especially in healthcare.
His first introduction to “research methods” was during an internship with a startup called Netradyne. On the topic of solving real impactful problems, Netradyne works on developing solutions in the intersection of autonomous driving and fleet management. This experience exposed Shekar to how the work described in research papers is the basis of a lot of innovation we see around us. This experience also furthered his interest in pursuing research in the field of computer vision.
I asked, “what is one big problem in the world that you’d love to solve?” Shekar said he would love to solve problems in the application of Artificial Intelligence (AI) to Healthcare. He had always been fascinated by the work that doctors do but never had the patience to work towards being a doctor. He has been working with Faraz Faghri and Professor Roy Campbell on solving problems in this area. There are many interesting projects being done in this field of research — everything from predicting the severity of Diabetic Retinopathy from retinal images to predicting how a person looks from gene sequencing data. With technology becoming more accessible to a large population, he hopes to make mobile healthcare technologies more accessible to people in need. He feels that any problem solved in the healthcare space has the potential to really affect people’s lives and that this kind of work helps him connect with humanity in general. It is extremely inspiring to see the impact this work can have on society at large.
In the future, Shekar hopes to make healthcare more accessible to people in developing countries. He hopes to ultimately be able to integrate his research into tools that can help people diagnose themselves. “I’d love to expand on this research and apply these methods to other open problems in the medical space.”
His favourite part about being in this line of work is having the opportunity to work with brilliant individuals who want to make a difference, the ability to work on open problems that are impactful, and the excitement he gets when he sees positive results after spending weeks in grueling agony trying to work towards a solution.
Shekar’s research is not without issues. In the machine learning space, everything revolves around data, and the unfortunate truth is that medical data is intrinsically messy. For example, the MRI images for different patients may have been taken using different imaging machines. Also there are numerous edge cases that you encounter when handling data in large volumes that makes this line of work a bit frustrating at times. Another major problem in the medical machine learning space is the shortage of data. When looking at deep learning models, especially in the case of image classification, we large datasets are required in order to achieve good results. The unavailability of data is a huge bottleneck and often hinders the progress of projects.
The main application of Shekar’s research would be to aid doctors in making medical decisions. However, there is a lot of room for improvement. Especially when discussing tools in the medical space, false negatives and positives can be extremely costly. As a result, they need to ensure that any tool that is being used in healthcare meets extremely high standards. In this line of research, it’s also important to understand how generalizable your solution is. For example, does it work well on all ethnicities? Machine learning tools are only as good as the data they have been trained on. As a result, it is important to analyze where your tool performs well and where it doesn’t.
Shekar describes himself as inquisitive, whimsical, and cooperative. Prior to this interview, he hadn’t thought of himself as a scientist. His favourite scientific journals include The Computer Vision Foundation, Neural Information Processing Systems, and Nature and the Journal of American Medical Association. Aside from reading, his hobbies include hiking, playing basketball, and snooker. He is often seen enjoying a good game of basketball on one of the various courts on campus, and always has a helping hand for friends. We can’t wait to see how you change the face of AI and healthcare!
Editor’s Note: In the original version of the story, Shekar’s last name was misspelled and since has been updated. We regret this error.