CNN based Chromosome Classification Architecture for Combined Dataset

Karyotyping is the procedure of examing the set of chromosomes from an individual by paring and ordering them. It features chromosome segmentation followed by classification. The resulted images can be used to diagnose genetic abnormalities, but they are usually manually obtained by cytologists, which is time-consuming and labour sensitive. Recently, researches have been carried out on chromosome image classification via Convolutional Neural Network (CNN) with significant progress. However, most studies focus on a single dataset for model evaluation, while chromosome images from different datasets are likely to show widely ranged characteristics. This research studies the model performance based on combining two public chromosome datasets, BioImLab and CIR. It proposes a new CNN network constructed by stacking a building block consisting of three convolutional layers followed by one residual layer using the Leaky-ReLU activation. Experiments demonstrate the efficacy of our proposed model, with 92.97% accuracy, 93.0% recall, and 92.98% precision using the combined dataset, outperforming the traditional ResNet50 and SENet50 networks.