Skip to main content

Minor criticism of publication 'Deep Neural Networks Improve Radiologists’ Performance in Breast Cancer Screening'


Publication Deep Neural Networks Improve Radiologists’Performance in Breast Cancer Screening is trying to detect cancer by using RESNETs. It does a better model by doing 4 RESNETs in parallel (instead of just single image). Yes, if cancer is found in CC and also MLO view, combining separate results should lead to improvement. What the model does not do the best way is how it combines the results:


As you can see they combine results after average pooling. This completely disregards positions of found cancer in MLO and CC image, and position of cancer and that it is located on both views at the same place (relatively due to different view) could lead to further improvement of the results (there should be correlation between MLO and CC finding [position of cancer] on the same breast).

From the publication: "Average pooling in all models is averaging globally across spatial dimensions in all feature maps."

Also, usage of average pooling instead of max pooling is little questionable, as cancer at initial state should be at small location, not all around. While average pooling for breast density makes sense (breast density is assigned by looking at whole breast area), for detecting cancer it does not make so much sense as signal should be coming from one area only - not multiples (unless the cancer is already way too spread).