I have a good friend who’s a pediatric orthopedic surgeon at Stanford (California), who is working on a project to predict the scoliosis (particularly to help primary care doctors identify kids who will likely need future treatment). My friend doesn’t have ML expertise, although he’s now tried to work with a few different groups. He suspects the current group he’s working with is using outdated techniques, and it struck me that maybe someone here in the fast.ai group would find this an interesting project to work on? They are looking to publish a paper if the results are strong (and more importantly will ideally want to turn this into a tool that primary care doctors can use).
A bit more about the project itself:
They are hoping to do this from 3D scans (from phone cameras) of the kids standing and bending over. My friend says he’d be able to identify kids easily from seeing these scans, someone who focuses even more could do it with only the scans of the kids standing, whereas primary care doctors (who don’t specialize in this at all) sometimes miss it entirely even from in person visits.
Currently they have a small amount of data (on the order of 200 scans paired with x-rays for kids who are affected. He thinks it should be easy to get a somewhat larger set of kids who would test negative, though obviously this has to be done with care around IRB and data privacy rules).
I work at UTRGV School of Medicine (UTHealth RGV) and also have an appointment as an adjunct faculty member in the Department of Computer Science. My job at UTRGV is creating the data and AI platform to empower doctors and researchers to do this type of work in a way that is compliant with legislation and data protection standards. This project seems exciting! I’d love to connect with your friend to discuss this project further. My goal with fast.ai is 1) to contribute to these types of efforts personally and 2) to do and understand the former really really well so that I can empower others with the technologies, platforms, and approaches such that I help proliferate AI in medicine at my institution and others. I will send you a DM with my contact information.
This sounds like a very doable project. I am definitely interested in hearing more.
Is the goal to predict whether there is scoliosis or not, or is it to layer visual aids to assist doctors in identifying scoliosis.
I suspect that simply predicting whether scoliosis exists or not should be very easy to do once the dataset is created. 200 x-rays might be enough - I would probably try pre-training on other publicly available x-ray datasets then fine-tuning to this scoliosis dataset to minimize the amount of data needed since there isn’t a ton of data. That said, people are often hesitant to use a tool that gives them an answer (ie positive/negative) if they can’t see it/explain ithe result (with good reason in a medical context). There’s also a matter of if a diagnosis by this model is acceptable evidence for billing based on the CPT code requirements of the insurance company. I suspect for it to be of widespread use it have to be “health insurance friendly”.
I think it’s more development work, but easier to implement a system that acts as a guide to assist doctors in the diagnosis rather than a system that only predicts the diagnosis. For example bounding boxes around the part(s) of the image that shows the abnormality best, possibly with a category label or other guidance for what could be the issue in that particular area. I think this is more practical immediately, but of course involves more work on the data science side. For example the dataset needs bounding boxes and categories. A object detection model is also slightly more complex. You also would likely want to consult some primary care doctors to see what they’d be comfortable using given their lack of specific expertise in this diagnosis. What information would they need to feel comfortable explaining the findings to the patient? How does that get built into the end product? Having a human in the loop like this also creates a path of correction when the model does issue a wrong prediction (and at some point it will, it’ll never be 100% accurate and neither will humans).
Is the goal just to predict the diagnosis? Or is the goal to give visual aides and guides to primary doctors so they are better able to make the diagnosis?
Anyways - I am interested and those are my initial thoughts on the topic. It’s all off the cuff, so I wouldn’t be surprised if I am misunderstanding some aspect of the problem. I hope it gives an idea of where my head is at and how I think about these kinds of problems.
A worthy project. In my childhood, my Scoliosis (and Kyphosis) were detected from a nurse school-visit program. I could imagine simple video recording follow a simple script of how to pose & bend could be effective in detecting a lot of missed cases out there.