Quantifying Similarity in Histopathology Images Using Pathology-specific Deep-learned Features

Interpretation of histopathology slides is often labor intensive and operator dependent. This work aims to support the pathologists by matching unknown slides with reference slides of similar characteristics (normal, benign, type of carcinoma, etc) from an annotated collection. We expand on prior efforts in this area by developing a set of pathology-specific deep-learned features for the similarity matching. The features were extracted from digitized breast histopathology slides using a convolutional autoencoder (CAE). For training, random patches of 256x256 pixel ROIs (tiles) were extracted from each slide and augmented with random affine transformations and perturbations in HSV color space to consider variability in orientation, magnification, staining and lighting conditions. Batch sizes of 128 were used in training for 2000 epochs. To assess the ability to detect similar images using the deep-learned features, we compared feature vectors of neighboring tiles of the same slide using a separate validation slide set. For inference, input tiles extracted from unknown slides are featurized with the encoder and a kNN search is performed on feature vectors derived from an annotated reference collection, returning similar tiles and corresponding slides for the pathologist.