Deep Metric Learning for Identification of Mitotic Patterns of HEp-2 Cell Images

Abstract

Automatic identification of mitotic type staining patterns in microscopy images is an important and challenging task, in computer-aided diagnosis (CAD) of autoimmune diseases. Such patterns are manifested on a HEp-2 based cell substrate and captured via Indirect immunoflourescence (IIF) based microscopy imaging technique. The present study proposes a deep metric learning methodology, in order to identify the mitotic staining patterns which are rather rare, among several other interphase patterns present in majority. Hence, the problem is framed as a mitotic v/s non-mitotic/interphase pattern classification problem. Here, the implemented network maps the input images into a latent space, in order to compare the distances between the samples, for class declaration, via a triplet-loss based framework. The framework yields good classification performance as max. 0.85 Matthews correlation coefficient in one case, with less false positive cases, when validated over a public dataset.