Combining Nakagami imaging and convolutional neural networks for breast lesion classification

The quantitative ultrasound (QUS) imaging provides additional information on tissue properties in comparison to standard ultrasonography. In this paper we use the Nakagami imaging to address the problem of breast lesion classification. Images of breast lesions may contain areas with calcifications or/and necrosis. These areas are visible on Nakagami maps and their presence make the classification more difficult; efficient texture features for classification are harder to estimate. As a remedy, we propose to use a convolutional neural network which automatically learn task dependent patterns from images. We train the network based on Nakagami maps of breast lesions.