Donate to Science & Enterprise

S&E on Mastodon

S&E on LinkedIn

S&E on Flipboard

Please share Science & Enterprise

$1.2M Challenge Seeking More Accurate Mammograms

Mammogram

(National Cancer Institute)

7 September 2016. A new challenge is seeking better computer models that interpret mammograms with fewer erroneous results, resulting in fewer unneeded medical tests. The crowd-sourced competition, conducted by Sage Bionetworks and Dream Challenges, has a total prize purse of $1.2 million, with submissions of models beginning on 4 October 2016.

The Digital Mammography Dream Challenge aims at producing more predictive algorithms for interpreting mammograms, with the goal of reducing the recall rate from mammography screening. Mammography is the most widely used method of breast cancer screening for women age 40 to 74, which has shown to be effective in reducing mortality. But mammography is far from perfect, with risks both of false negatives, where the exam misses a breast tumor, or false positives that indicate a tumor when in fact no cancer is present.

The challenge is aimed at false positives, where National Cancer Institute estimates on average 1 in 10 women having a mammogram will be called back for more testing. In general, that percentage is double the actual rate of 1 in 20 having breast cancer. NCI says as many as half of the women screened annually with mammograms will encounter false positives, of which 7 to 17 percent will have unnecessary biopsies. This evidence and others led the United States Preventive Services Task Force and American Cancer Society to recently revise their guidelines on mammograms.

The challenge is seeking new algorithms for interpreting mammography images from inside and outside the bioinformatics community. New algorithms should reduce the rate of false positives while still detecting real breast tumors. The algorithms should also indicate a high or low likelihood of cancer from the image, and whether further testing is warranted. In addition, the models can indicate if customized screening routines are needed, and if the woman should increase or decrease the frequency of mammograms.

Organizers of the challenge anticipate the use of machine or deep learning to train the models in interpreting mammogram images. For this purpose, Group Health Cooperative in Seattle and the Mount Sinai medical school in New York are making available more than 640,000 anonymous mammogram images from more than 86,000 patients, along with corresponding clinical data.

Individuals may take part in the competition or form teams, but in each case all participants must register. Because of the size and complexity of the data, participants cannot download the data sets, but are asked to submit their models and algorithms for training with the data. Scores from the algorithms will be posted on leaderboards in 3 rounds of submissions, each 4 weeks long, beginning on 4 October. Each team or individual participant can submit 3 models in each of the 3 rounds. A validation period begins on 6 February 2017, where participants enter their final models, with final scoring beginning on 20 February and ending on 13 March 2017.

In the leaderboard phase, $20,000 in prizes will be awarded for the top 3 models in each of the rounds. Prizes of $100,000 each plan to be awarded for the top models in each of 2 sub-challenges during the competition. Top performers will then be invited to take part in the community, or final competition, with a total of $1 million expected to be awarded to the top performing individuals or teams.

The Digital Mammography Dream Challenge is also part of the Coding4Cancer program that helps sponsor these competitions. A similar challenge is expected to take place in 2017 aimed at improving lung cancer screening techniques.

Read more:

*     *     *

Comments are closed.