presents unique challenges compared to traditional natural image processing. Unlike natural images, where multiple objects are present, mammograms often contain only a single focal point, even in cancerous cases. Consequently, applying models designed for natural images to medical imaging is not advisable. “If you take a model designed for natural images, every image has an object,” Nhi continues. “Every image can be used to train the model. In medical imaging, many images are negative, so there’s no object at all. How can you use these images to train the model?” Also, breast cancer screening typically involves capturing two images of each breast, providing complementary views of potential findings. A finding can look suspicious on one view but normal on the other. These views must be considered together to make accurate decisions. “Our director, the last author, is very interested in multi-view reasoning, so he came up with this idea of working with two different views,” Nhi recalls. “To make multi-view reasoning work, we must limit how many boxes or proposals we have on each view. That’s when I started to look into the sparsity. At this point, Dan joined and did a lot of work on multi-instance learning.” As each malignant image typically contains only one finding, M&M uses a sparse detector with a set of sparse proposals, limiting the number of false positives. Dense detectors, where you have many dense proposals and anchors, tend to generate many false positives. Furthermore, introducing a multiview attention module facilitates reasoning between the two views, enhancing the accuracy of detection, and multi-instance learning allows the team to harness seven times 13 DAILY MICCAI Tuesday M&M withNhi andDan
RkJQdWJsaXNoZXIy NTc3NzU=