DoMore! • Deep learning & Convolutional neural networks!

DoMore! • Deep learning & Convolutional neural networks!


Automation is driving changes in many
industries: transportation, education, health care, and most other fields are
expected to undergo tangible disruption in the near future. One of the
technologies behind this development is machine learning. Machine learning
describes a family of methods that solve a certain task by learning from
experience. One of these methods is neural networks. At the Institute for
cancer genetics and informatics we employ a combination of deep neural
networks and mathematical concepts often called deep learning deep refers to the
multiple layers in the network deep learning is central to the development
of automated services such as driverless cars digital assistants object
recognition and automatic translators in the do-more project at AI CGI we use
deep learning in cancer research applications for example in the
automation of delineating and grading tumors this means identifying malignant
tumor areas and determining the prognosis a major tasks carried s by the
pathologist the specialized physician who diagnoses cancer for this
demonstration we will treat the image as a whole
a colored digital image is composed of three primary color channels carrying
intensity information in a grayscale format in the analysis of these images
the computer sorts these pixels into a matrix of numbers for each pixel the
computer assigns a numerical value from zero to one based on a grayscale where
black represents an intensity of zero and white represents full intensity one
to extract image features filters are used a filter is a matrix of numbers
which when applied to the image matrix extracts different features to make this
process visible we have scaled up the pixels we generate our own filters as
part of the training process these filters have the specific purpose of
identifying tumor features a layer of convolved images is generated by running
filters across the grid convolutional neural networks consists of many such
layers where each image is extracting different features from the previous
layer we then use local operations to downscale these images to produce a
so-called pooling layer this is done to aggregate the features and reduce
dimensionality a second convolved layer is generated by running a different set
of filters over the pooled images in this case 16 filters in total the
pattern of convolving and pooling continues until we have generated
sufficient feature information network configurations can range from just a few
to thousands of layers leading to many millions of trainable parameters we
consolidate the final layers of feature images which results in values
representing the probability of the patient’s outcome the difference between
the predicted output and the desired output is measured using a mathematical
function often called a loss function this calculation provides a point of
orientation that guides the necessary adjustments to the filter values we then
use an iterative optimization method to minimize this loss function
it uses a method called back propagation to compute the values needed to adjust
the filter parameters in the network this is done by sending new images
through the neural network and adjusting the initial filters until the output
image is approximating the ground truth well enough this process is called
training deep learning utilizes massive amounts of labeled data commonly known
as big data in our case this means the image of the patient’s cancerous tissue
along with their follow-up information for colorectal cancer we have over 4,000
cases with more than five years of follow-up information for the patient
group data sets of this size with such extensive clinical follow-up are
incredibly rare giving us a unique position when applying deep learning to
pathology

4 Comments

Leave a Reply

Your email address will not be published. Required fields are marked *