Ensemble of Unsupervised Learned Image Representations Based On Variational Autoencoders for Lung Adenocarcinoma Subtype Differentiation

Lung adenocarcinoma is a type of non-small cell lung cancer that accounts for about 40% of all lung cancers, which is divided into different molecular and histological subtypes associated with particular prognosis and treatment. Pathologists stratify for diagnosis mainly by its histo-morphological visual features and patterns, which tends to be challenging because of the nature of lung tissue, a mixture of histologically complex patterns and not having a specialized grading system. Here, an unsupervised computational approach based on an ensemble of tissue-specialized variational autoencoders, which were trained per histopathology subtype, to build an unsupervised embedded tissue-image representation. This representation was used to train a Random Forest classifier of three lung adenocarcinoma histology subtypes (lepidic, papillary and solid), and a 2D-visually interpretable projection from the learned embedded representation. Experimental results achieve an average F-score of 0.72 ± 0.05 in the test dataset and a well-separated 2D visual mapping of tissue subtypes.