Deep Learning for Medical Image Analysis, specially breast cancer diagnosis and prognosis

Dnr:

SNIC 2017/3-56

Type:

SNAC Small

Principal Investigator:

Hossein Azizpour

Affiliation:

Kungliga Tekniska högskolan

Start Date:

2017-05-22

End Date:

2018-06-01

Primary Classification:

10207: Datorseende och robotik (autonoma system)

Webpage:

Allocation

Abstract

With the advent of high-end GPUs and the new techniques for training deep neural networks, the fields of machine learning in general and computer vision in particular have witnessed an unprecedented leap of performance. The field of medical image analysis provides many interesting and important tasks where deep networks can be applied. The purpose of this project is to try various recent advances in the field of deep learning on different diagnostic and prognostic tasks. Within this project the main focus will be on analyzing mammograms from screening for breast cancers for diagnostic purposes as well as analyzing whole tissue slides from histopathology of sampled breast tissue for prognostic and treatment purposes. The size of images in medical image analysis are usually much larger than general computer vision. A typical mammogram image is 5k by 5k pixels and a typical whole slide tissue scan can be as large as 100k x 100k pixels. This raises new challenges and makes the usage of many GPUs concurrently very cruicial. This is a small project and is basically to see the feasibility of doing our analysis on the available machines. If successful we will apply for medium/large allocations.