Investigation of KimiaNet's and DenseNet's deep features in lung cancer subtypes

Abstract

Deep neural networks (DNN) have extended applications in the _eld of digital pathology. One of which is to act as feature extractors for content-based image retrieval (CBIR) systems. Therefore, it is necessary to investigate how these deep features work and attribute these features to histologic patterns. This study showed that the median of deep feature value could serve as a simple yet efficient representation of whole slide images (WSI). Through exploring deep features of lung cancer, it was discovered that some of these deep features have strong correlations with either lung adenocarcinoma (LUAD) or lung squamous carcinoma (LUSC). A deep feature-specific visualization technique was proposed for analyzing deep features at WSI-level. These prominent deep features were also generalizable to cancers of other organs, namely kidney and brain. However, deep features did not exhibit a potential for cancer grade and somatic mutation state classification.