Using CUDA and Machine Learning to Detect Colon Cancer

As part of the GlaS@MICCAI2015 colon gland segmentation challenge, a team of researchers introduced a machine learning-based algorithm to segment glands in tissue of benign and malignant colorectal cancer.

The variability of glandular structures in biological tissue poses a challenge to automated analysis of histopathology slides. It has become a key requirement to  quantitative morphology assessment and supporting cancer grading.

Using GPUs to detect colon cancer
Qualitative segmentation results on images: Segmentation (blue outline) and ground truth (green outline), false negative pixels are cyan, and false positive pixels are yellow

Using GPUs, CUDA, and Pylearn2 — a machine learning library built on top of Theano — the team trained their two deep convolution neural networks on a set of 125,000 images and achieved a classification accuracy of 98% and 94%, making use of the inherent capability of the system to distinguish between benign and malignant tissue.

In related news, the NVIDIA Foundation recently awarded $200,000 to a team of researchers from the University of Toronto for their GPU-accelerated cancer research by developing a “genetic interpretation engine” – a deep learning method for identifying cancer-causing mutations.

Read the research paper >>

3 thoughts on “Using CUDA and Machine Learning to Detect Colon Cancer

  1. Hao Chen on December 3, 2015 at 2:06 pm said:

    We are the winner of Glas@MICCAI2015, more details about the method will be released soon

    • brad nemire on December 3, 2015 at 2:09 pm said:

      Hi Hao – Did your team use NVIDIA GPUs? If so, please comment here with the details. Congrats!

      • Hao on December 3, 2015 at 2:14 pm said:

        Hi Brad, thanks.
        Yes, we used the wonderful NVIDIA GPU TITAN X in our implementation of deep convolutional neural network. Different from the team mentioned above working on the classification task, we focused on the challenging segmentation of gland in the pixel-level, ultimately we want the system to facilitate the automatic diagnosis.