Extended ResNet and Label Feature Vector Based Chromosome Classification

Human chromosome classification is essential to the clinical diagnosis of cytogenetical diseases such as genetic disorders and cancer. This process, however, is time-consuming and requires specialist knowledge. Considerable efforts have been made to automat the process. Recently, methods based on Convolutional Neural Networks achieved state-of-the-art results on the chromosome classification task. Many studies used karyotype images in performance evaluation, few studies have reported the results of human chromosome classification on microscopical images. This paper proposes a novel method to classify single chromosome images into one of 24 types. In the proposed method an extended ResNet was first devised to extract features of single chromosome images. A label feature vector was then extracted for each of 24 chromosome types based on a validation dataset. Hausdorff distance between feature vector of an input image and each of 24 label feature vectors were calculated, and the label feature vector that has minimum hausdorff distance to the feature vector of the input image was selected as the potential label of the input image. To finally allocate the single chromosomes from a same microscopical image into one of 24 types, a Label Redistribution strategy was used to shrink the label space and to increase the efficiency of chromosome classification. Experiments were implemented with 90,624 single chromosome images, 644 of which were randomly picked to form a testing set in advance. The classification accuracy on microscopical images using our proposed method achieved an accuracy of 94.72%.