Continuously learning AI pathologist: An AI powered smart microscope that can automatically scan different biological samples Tathagato Rai Dastidar, Co-founder & Chief Scientific Officer, SigTuple
The Wonder Tool! The clinical microscope Still a gold standard for detecting various types of abnormalities in clinical laboratories
… and the Wonder Person! The clinical pathologist
But there aren’t enough ... Some numbers: HOW MANY OF YOU ARE THERE? In India, - 1.2 Billion people - 20,000 pathologists! NOT ENOUGH
Introducing ... The continuously learning AI pathologist
Transformation over a year ...
How it works
Seeing is believing! Enabling remote review and collaboration
Multi purpose ... Blood Urine Semen … and many more in the pipeline
Continuous learning Re-training New data Extensive validation Deployment Unidentified cells
Continuous learning ● Predictions model become better, without any changes in the hardware ● Predictors for new disease conditions made available
Where we are ... ● Several clinical studies with leading laboratories ● High correlation of statistical indices with existing state-of-the-art ● Adept at finding rarer cells which a pathologist or existing machines typically miss ○ Proven across multiple studies
Where all does AI play a role?
Scanning ● Impossible to capture entire slide (75mm X 25mm) ○ Location and size of “analyzable area” uncertain ○ Use case dependent definition of “good 0.5mm area” diameter ● Intelligent way to figure out “promising” areas ● Machine learned models to identify “good images” and “good movement directions”
Classification Normal, elliptical, fragmented ... RBC WBC Neutrophil, lymphocyte, monocyte ... Normal, large ... Platelet ● Images are 4k X 3k, with thousands of objects ○ Cannot be classified as a whole ● General approach: Localization followed by classification ● Performance a major bottleneck for CNN based localization
Volume estimation 3D characteristic estimation from 2D images Red blood cells under a What they actually look like microscope
Preventive maintenance Dear customer, your device is likely to develop Machine motor problems operation logs in the next 48 hours ... AI/ML based analysis
Thank you! Questions?
Recommend
More recommend
