Multi-View Hypercomplex Learning for Breast Cancer Screening

Traditionally, deep learning methods for breast cancer classification performa single-view analysis. However, radiologists simultaneously analyze all fourviews that compose a mammography exam, owing to the correlations contained inmammography views, which present crucial information for identifying tumors. Inlight of this, some studies have started to propose multi-view methods.Nevertheless, in such existing architectures, mammogram views are processed asindependent images by separate convolutional branches, thus losing correlationsamong them. To overcome such limitations, in this paper, we propose amethodological approach for multi-view breast cancer classification based onparameterized hypercomplex neural networks. Thanks to hypercomplex algebraproperties, our networks are able to model, and thus leverage, existingcorrelations between the different views that comprise a mammogram, thusmimicking the reading process performed by clinicians. This happens becausehypercomplex networks capture both global properties, as standard neuralmodels, as well as local relations, i.e., inter-view correlations, whichreal-valued networks fail at modeling. We define architectures designed toprocess two-view exams, namely PHResNets, and four-view exams, i.e., PHYSEnetand PHYBOnet. Through an extensive experimental evaluation conducted withpublicly available datasets, we demonstrate that our proposed models clearlyoutperform real-valued counterparts and state-of-the-art methods, proving thatbreast cancer classification benefits from the proposed multi-viewarchitectures. We also assess the method generalizability beyond mammogramanalysis by considering different benchmarks, as well as a finer-scaled tasksuch as segmentation. Full code and pretrained models for completereproducibility of our experiments are freely available athttps://github.com/ispamm/PHBreast.