Triplet loss for Chromosome Classification

Abstract

The analysis of chromosomes, known as karyotyping, is essential in diagnosing various human genetic disorders and chromosomal aberrations. It can detect a variety of genetic diseases and provide a deeper insight into the human body. However, the process of manual karyotyping is highly time-consuming and requires accomplished professionals with a deep understanding in the field. An automated process is thus highly desirable to assist cytogeneticists and mitigate the cognitive load procured during karyotyping. With that intention, a similarity learning approach is proposed in this paper using ‘Triplet Loss’ for procuring high-dimensional embeddings. The Offline Triplet Loss, Semi-Hard Online mining, and associated hyperparameters are thoroughly tested and explored, and the obtained embeddings are used to classify the images into their respective chromosome classes and Denver groups. A comparative analysis on various embedding-classifying algorithms such as Multi-Layer Perceptron (MLP) and Nearest Neighbours is also demonstrated in this paper, along with experiments on associated distance metrics. The proposed methodologies deliver a superlative performance when compared to a baseline Convolutional Neural Network (CNN), on a publicly available chromosome classification dataset.