|
This study focuses on the problem of finding pathologies on mammogram screenings using computer vision-based systems. For that we use deep neural networks and use open-source datasets for our models training for finding segments of pathologies on screenings. These datasets are also used for further analyses of different supervised training techniques, functions and losses to achieve best results for generalization of data even when its’ domain changes. We don’t use screenings with pathologies during training while trying to achieve the results where our network would be able to still recognize these pathologies and embeddings (feature maps) extracted at certain layers of our network would allow to group up the samples where these pathologies are present by calculating either cosine or normalized L2 distances between samples. First, we train our network to recognize normal mammogram screenings and anything that is different. We create a vocabulary of known classes using extracted embeddings of samples from training subset to obtain centroids of embeddings for 2 separate classes – benign and malignant tumors. We then use a method known as CAM (Class Activation Mapping) along with Selective Search algorithm for segmentation task. In our work we use mainly dataset CBIS-DDSM [1] for validation which consists of a few hundreds of annotated images of digitalized mammogram screenings. The entire pipeline of the system has the following steps: first, extracting the embeddings from training subset and clustering them into vocabularies for classes of benign and malignant tumors; then, the clusters can be analyzed and regrouped manually depending on the situation to create sub-classes such as small or big benign tumors; afterwards, we use these existing vocabularies for classification task to detect screenings with tumors; and to obtain segments of the detected pathologies CAM along with Selective Search with carefully selected parameters is used.
|
