SPIE Membership Get updates from SPIE Newsroom
  • Newsroom Home
  • Astronomy
  • Biomedical Optics & Medical Imaging
  • Defense & Security
  • Electronic Imaging & Signal Processing
  • Illumination & Displays
  • Lasers & Sources
  • Micro/Nano Lithography
  • Nanotechnology
  • Optical Design & Engineering
  • Optoelectronics & Communications
  • Remote Sensing
  • Sensing & Measurement
  • Solar & Alternative Energy
  • Sign up for Newsroom E-Alerts
  • Information for:
    Advertisers
SPIE Photonics West 2019 | Call for Papers

2018 SPIE Optics + Photonics | Register Today

SPIE Journals OPEN ACCESS

SPIE PRESS

Print PageEmail Page

Biomedical Optics & Medical Imaging

Martin Stumpe: Advancing Cancer Diagnostics with Deep Learning

A keynote presentation from SPIE MEdical Imaging 2018.

1 March 2018, SPIE Newsroom. DOI: 10.1117/2.3201802.02

Martin Stumpe, Google Research (USA)Rendering cancer diagnoses from biopsy slides involves challenging tasks for pathologists, such as detecting micro metastases in tissue biopsies, or distinguishing tumors from benign tissue that can look deceivingly similar. These tasks are typically very difficult for humans, and, consequently, over- and under-diagnoses are not uncommon, resulting in non-optimal treatment.

Algorithmic approaches for pathology, on the other hand, face their own set of challenges in the form of gigapixel images, proprietary data formats, and low availability of digitized images let alone high quality labels. However, advances in deep learning, access to cloud based storage, and the recent FDA approval of the first whole slide image scanner for primary diagnosis have set the stage for a new era of digital pathology.

In this keynote presentation, Martin Stumpe of Google Research, discusses the potential of deep learning to improve the accuracy and availability of cancer diagnostics, and highlights some recent advances towards that goal.

"Deep learning, or DL, is a paradigm shift in machine learning," says Stumpe. "DL is a modern reincarnation of artificial neural networks -- collection of simple trainable mathematical units, organized in layers, that work together to solve complicated tasks."

But how good is DL in practice? Stumpe notes the ImageNet Large Scale Visual Recognition Challenge - which evaluates algorithms for object detection and image classification at large scale.. Since 2011, DL has won this challenge, with significant accuracy increases. DL can generally outperform generalist humans, but not experts - such as telling a husky from Alaskan sled dog.

Deep Learning_seeing in layers

Stumpe points out that one criticism is that "DL is like a black box, we don't really understand what's going on; we don't know what the predictions are based on; but I think it's possible to put the explainability in to DL."

Deep Learning_opening the black box

He adds that DL is not replacing pathologists, but it is a tool that's good at detecting metastases. Current models are getting good enough to be useful as assistants for pathologists. What's still being worked on are interoperability and user interface.

Image quality is also an issue as "parts of image can't be out of focus." Scanners are getting better, says Stumpe, but are still limited and need to detect artifacts.

"The most important thing to remember," he says, "don't trust ground truth from a single source."

Martin Stumpe leads the Pathology project at Google Research. Before that, he worked on the Google Street View project for automatically building maps using machine learning. Prior to joining Google, Martin worked on NASA's Kepler Mission to detect extrasolar planets, and worked on AnTracks, an interactive computer vision software for detection and tracking of objects in videos.

Stumpe received a PhD in physics from the Max-Planck-Institute in Goettingen (Germany), researching the molecular mechanisms of protein folding and stability, a topic  he continued during his postdoc research at Stanford University (USA).

Multimedia presentations from SPIE Medical Imaging 2018