Institute of Information Science, Academia Sinica

Events

Print

Press Ctrl+P to print from browser

TIGP (BIO) – Advancing Pathology with AI-Powered Digital Workflow

:::

TIGP (BIO) – Advancing Pathology with AI-Powered Digital Workflow

  • LecturerDr. Chao-Yuan (Joe) Yeh (CTO and Co-founder of aetherAI)
    Host: TIGP- Bioinformatics Program
  • Time2019-12-10 (Tue.) 14:00 ~ 16:00
  • LocationAuditorium107 at IIS new Building
Abstract

The core process of histopathology slide reading has largelyremained the same in the past one hundred years. Pathologists examine thinsections of tissue specimen through the binoculars of a microscope. Despite readilyavailable software and hardware solutions, the adoption of digital pathologyworkflow has been very slow. One of the main reasons is that return oninvestment is difficult to assess. The advent of AI-powered image analysis hasdramatically changed this perception and created unprecedented interest indigital pathology. Deep neural networks have been shown to outperform humanexperts in several pathology reading tasks, such as detection of metastaticcancer cells in lymph nodes. In this talk, I’ll share some of the challengingdigital pathology AI cases we’ve been working on. First I’ll discuss how we applydeep learning to the task of differential counting of bone marrow smear images.This task has been regarded as highly challenging even for human experts.Through curation of an extremely large dataset and iterative label correction,we’re able to train a deep neural network that performs bone marrow smear differentialcounting on par with human experts. Second, I’ll discuss how we overcome thelimitation of GPU memory on digital pathology AI. The extreme size of digitalwhole slide images necessitates patch-based methods when it comes to deeplearning because of GPU memory constraints. While mechanism such as nVIDIAunified memory exists and allows GPU to use system memory for its compute, theslowness of data migration through hardware interfaces makes it impractical.I’ll discuss how we optimized compute graph through group execution and groupprefetch and achieved over 200% improvement in throughput when we trainedconvolutional neural network on input images as large as 10000*10000 pixels.I’ll further share results on using this approach to train deep neural networkson whole slide images with only sign-off diagnosis but no detailed annotations.We’ll show evidence that this approach is more data-efficient when comparedwith methods such as multiple instance learning. I’ll finish the talk with someof the directions we will be exploring when we try to overcome the challengesof digital pathology AI.

BIO

 

https://tigpbp.iis.sinica.edu.tw/tigpbio/CV_Joe%20Yeh_2019.pdf