Coverage by Bhat Dittakavi of Variance.AI on Utsah talk, Preventive Healthcare AI, at IIITH by Yogesh Bathina on 4th March 2017
Medical image analysis (wellness and preventive care)
I was into Machine Learning and Computer Vision. Now in Transin.in, Uber for trucks. We are a 100 crore company and Third largest in India.
Why are we limiting use of technology for diagnostic reasons only? Why not we do it for routine wellness check?
The pot belly (outer body part) is subcutaneous (under the skin). Visceral fat is the fat between organs that is dangerous that leads to many diseases. Fat is distributed all around the body. What if the toning happened but the visceral fat is still there. How the fat is distributed around the organs really matters. Around the organs and bones is the key.
Gym: Fitness, prevention and weight loss
Diagnostic imaging: Detects what’s wrong inside.
Can we marry these two? that AI our idea.
CT is radiation. What about MRI? It isn’t.
There is one studio in Apollo that charges Rs.80000 for doing the same.
You worked out 15 days in a gym and got an MRI scan done. Radiologist advises you to do more abs with so and so supplements and diet to reduce the fat around an organ area.
This has three parts..
1) Clinical research (where is fat)
2) Image analysis
3) Viable business model
MRI is just exciting molecules. Dixon Imaging is one way to detect fat. We will be able to isolate the scan into water, in-charge, out-phase and fat. Fat area is of our interest.
Lab test data
Say you got 100 images of face and get mean face. Similarly we do 100 images of organs and we get the healthy organ image.
We use random forest also to detect multiple parts around the body.
Take the mean model in a mesh and iterate through laplasian model as we move this mesh until it exactly matches the healthy organ. Normal and gradient search are aligned at one particular level.
The business model
We set up in Vedanta Medicity. We could scan the entire body (Fast scan) under a minute where as the typical MRI scans take 40 minutes.
MRI could cost Rs.30,000 to 40,000. It costs between Rs. 2000 to Rs. 3000 per scan. The solution requires a combination of experts.
-Radiologists (analyse the scan)
-Cardiologists (analyse the fat around heart)
-Endocrinologist (diabetics type)
If there is an MRI scan in the vicinity of a fitness center, we are good.
Quantification of what fat is present, Over time we can use multi-time registration models to assess risk.
Big data context
We have tons of anonymous data. Take we got. Take the case of heart attack. We have two options. Open heart or stunting.
Heart is blocked, open heart surgery.
It takes lakhs for stunting. No standard. If we have a big data infrastructure, we will be able to ask questions on how many people like me got stunt recommended (instead of taking a chance, do a data modeling).
HIS (Health Information System: Patient info)
We plug all these data streams into the data lake so we can do analytics. We can use this big data to compare the data with the others.
Pre-labeled algorithm with samples is faster.
We filed a patent. It is going great through clinical approval. It takes 7-8 years to get into production.
Q) How did you arrive on random forest?
We started with Adaboost and then to Random Forest. It worked with initialization well then is faster.
Q) Engineering hurdles?
Funding! Tough to get doctors in board. Implementation side we got a detection and segmentation but we have other problems in ML to solve.
Q) Why approvals?
Ethics committee of the hospital has to approve even if you collect anonymous data. You can use it as research tool but can’t multiply as business.
Q) Why 7 years?
To study. What part of algorithms we need to change? We need to identify control group. Study like enough to make a product out of it.
Even iron shows up as fat. We need to do some MRI corrections for this.
Data from Indians is different from the western world. Indian data is a scarcity.
Q) Where do you get revenues?
We are part of the Siemens corporate funding. We compete for prizes.
Q) Are models gender specific?
We work with bariatric surgeries. We aggregated 24000 trucks. We are inner-city. 100 crores in 1.5 years.