Deep Learning Helps Spot Dense Breast Tissue in X‑Ray Images

MammoDenseSegNet is a new computer program that looks at breast X‑ray pictures, called mammograms, to find the parts of the breast that are very dense. Dense tissue is a sign that a woman might be at higher risk for cancer, so doctors need to see exactly how much of it is there. The job is hard because the pictures can look very different depending on where they were taken and what machine was used. To make the job easier, the designers built two special tools into their program. First, an “adaptive dual attention” part of the network can focus on important spots in the image no matter where they are. It does this by looking at how every part of the picture relates to every other part, both in space and in color. Second, a “multi‑kernel receptive field” piece makes the network look at many different sizes of patterns all at once. This helps it pick up clues that are small or large, giving a fuller picture of the tissue.The team also added a clever way to teach the program, called multi‑scale dice loss with deep supervision. This method tells the computer to pay attention at several stages of its own processing, which makes it more reliable. To test the new model, researchers used three sets of images: two public collections (VinDR‑Mammo and EMBED) and one private set. In total they examined 1, 499 pictures from 606 women that showed a wide range of breast densities and some image glitches. They used statistical tools to make sure the results were fair, even when many pictures came from the same woman. MammoDenseSegNet performed very well in every situation, with recall scores between 0. 64 and 0. 90 and dice scores from 0. 63 to 0. 91. These numbers were much better than those from a popular older program that was built on the VGG16 network, which had recall ranging from 0. 04 to 0. 91 and dice from 0. 06 to 0. 82. The biggest advantage of the new model was when the breast tissue was not very dense; the old program almost never found anything useful there, while MammoDenseSegNet still gave useful results.