Miki 1 Shouhei Hanaoka 1 Yukihiro Nomura 2 Takahiro Nakao 3 Saori Koshino 3 Takeharu Yoshikawa 3 Naoto Hayashi 3 Osamu Abe 4 1 Department of Radiology, The University of Tokyo Hospital 2 Center for Frontier Medical Engineering, Chiba University 3 Department of Computational Diagnostic Radiology and Preventive Medicine, The University of Tokyo 4 Graduate School of Medicine, The University of Tokyo
only using virtually created X-ray images (thus, the training dataset is free from privacy concerns) LUPIN-AI Learned Using Purely Imaginary Nodules Here’s Our AI (= Your Rival)
• can estimate probability density function (PDF) of the given datasets (e.g., human chest X-ray images) • can provide a latent vector expression of a given sample (e.g., an X-ray image) • is not a method for dimension reduction; the length of the latent vector is the same as the input vector (image) • can be used to interpolate two different images (and more) https://openai.com/blog/glow/
lung / heart diseases • pneumonia, lung cancer, heart failure, etc… • Computer-assisted diagnosis of chest X- ray images has to handle a wide range of diseases
Estimate the PDF of normal human chest X-ray images • 2nd step: Calculate the “normal sample” hyperplane • 3rd step: Calculate the latent vector expression of a given abnormal image • 4th step: Move the given abnormal image so that the point will go away from the original point – “abnormality emphasized image” “normal sample” hyperplane (in the latent vector space)
“normal sample” hyperplane (in the latent vector space) • 1st step: Estimate the PDF of normal human chest X-ray images • 2nd step: Calculate the “normal sample” hyperplane • 3rd step: Calculate the latent vector expression of a given abnormal image • 4th step: Move the given abnormal image so that the point will go away from the original point – “abnormality emphasized image”
given abnormal image “normal sample” hyperplane (in the latent vector space) • 1st step: Estimate the PDF of normal human chest X-ray images • 2nd step: Calculate the “normal sample” hyperplane • 3rd step: Calculate the latent vector expression of a given abnormal image • 4th step: Move the given abnormal image so that the point will move away from the normal hyperplane – “abnormality emphasized” image
given X-ray image is presented. • It can emphasize arbitrary types of abnormalities, thus we expect that it can be used in daily routine X-ray image diagnosis tasks.
& Gal, Y. What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision? in Advances in Neural Information Processing Systems 30 (eds. Guyon, I. et al.) 5574–5584 (Curran Associates, Inc., 2017).