Efficient Type and Polarity Classification of Chromosome Images using CNNs: a Primary Evaluation on Multiple Datasets

Abstract

Karyotyping is critical for screening genetic diseases in an early stage. However, the manual karyotyping process is a labor-intensive task. This paper focuses on automatic chromosome image classification, which is a step in karyotyping. We propose a convolutional neural network architecture for efficient type and polarity classification of the chromosome image, namely ETPC, which is leveraged from the EfficientNet family’s development. The ETPC’s classifier with a weighted classification loss function are designed for efficiency in the training process. We perform our experiment with two training scenarios on four clinical datasets. The experiment on a dataset of 28,225 original chromosome images demonstrates that the proposed network achieved comparable results, with an accuracy of 95.3% for the type classification and 99% for the polarity classification, while having a significantly smaller number of parameters than state-of-the-art methods. Furthermore, the experiment on multiple datasets shows that the proposed network can dramatically improve the classification accuracy on new datasets with a small amount of fine-tuning data.