Share this post:
One of the breast cancer samples from the Tumor Proliferation Assessment Challenge 2016 training data set.
Deep learning and neural networks are making significant progress in identifying cancer mitosis.
A critical step in the diagnosis of cancer is the analysis of a patient’s biopsy tissue sample, which sometimes can be as small as a pinhead. Even with such a small sample, pathologists can test for the absence or presence of tumor cells to provide important information pertaining to the course of treatment to doctors.
To analyze the samples, pathologists typically stain the tissue sample with liquid re-agents. The intensity and distribution of the color stain classify and determine the extent of the disease.
The stained tissue samples are then studied under a microscope, which can take long hours, particularly when reviewing hundreds of images a day. Due to this Big Data overload eye fatigue is common, leading to diagnosis errors.
With recent developments in image recognition and deep learning, computer scientists believe they may have an answer and pathologists are eager for the help. To check on the readiness of the technology healthcare professionals around the world are organizing hackathons and challenges.
One such challenge took place a few weeks ago with support from the University Medical Center Utrecht, Eindhoven University of Technology, Beth Israel Deaconess Medical Center and Harvard Medical School. The organizations hosted the Tumor Proliferation Assessment Challenge in Athens, Greece as part of the MICCAI 2016 conference. 159 teams from around the world downloaded a training data set of 500 breast cancer images with image resolutions, in some cases, exceeding 50,000 x 50,000 pixels. The challenge wasn’t easy and in the end only 14 teams submitted results.
One of them included scientists from IBM’s research labs in Switzerland and Brazil. The international team of IBM researchers, who are French, Hungarian, Brazilian and Greek participated in the “mitosis detection task using automatic methods” challenge. Joining the competition several months late led to a busy summer, but it was worth it, as the team placed second, missing top honors by only 0.004 points.
IBM scientists, Erwan Zerhouni, Maria Gabrani and David Lanyi, are tackling cancer with deep learning and neural networks.
“The task was spot the mitosis, which is incredibly tricky for a human, let alone a computer,” said David Lanyi, who is studying deep learning at ETH Zurich when he is not at IBM.
“We started training our deep learning algorithm using neural networks in July to recognize the difference between a positive and a negative tissue sample and after using a few tricks, it learned rather quickly.”
“Five years ago it would have been impossible to even consider this task,” said Erwan Zerhouni who joined IBM for his master thesis.
“Currently, it takes one hour to diagnosis a 5600 x 5600 image, but as we develop this further we should reach under a minute, or even 20 seconds. And it can be trained for any type of cancer.”
Matheus Viana, who participated in the challenge from IBM’s lab in Brazil, is already thinking about the future.
“We envision combining the current deep learning techniques from the MICCAI challenge together with omics data (genomics and proteomics) for precision medicine and better prediction of patients outcome.”
The team plans to share their work with scientists at IBM’s Haifa lab, which are also developing imaging analytics for breast cancer.
But cancer is just one disease that can benefit from IBM Research’s work in medical imaging. IBM Fellow and medical imaging expert Dr. Tanveer Syeda-Mahmood, also heard about the results and expects to work with the team to see how it can be incorporated into her Medical Sieve research, which aids radiologists and cardiology assistants in diagnosing heart disease. These medical professionals often suffer the same challenges as pathologists when it comes to eye fatigue. Syeda-Mahmood’s work will be featured at next week’s Radiological Society of North America’s Annual Conference.