Automated system for chromosome karyotyping to recognize the most common numerical abnormalities using deep learning

Chromosome analysis is an essential task in a cytogenetics lab, where cytogeneticists can diagnose whether there are abnormalities or not. Karyotyping is a standard technique in chromosome analysis that classifies metaphase image to 24 chromosome classes. The main two categories of chromosome abnormalities are structural abnormalities that are changing in the structure of chromosomes and numerical abnormalities which include different types like monosomy (missing one chromosome) or trisomy (extra copy of the chromosome). Manual karyotyping is complex and requires high domain expertise, as it takes an amount of time. With these motivations, in this research, we automate karyotyping by borrowing the latest ideas from deep learning and recognize the most common numerical abnormalities (monosomy and trisomy) on a dataset contains 147 non-overlapped metaphase images collected from Center of Excellence in Genomic Medicine Research at King Abdulaziz University. The metaphase images go through three stages in our study. The first one is individual chromosomes detection using YOLOv2 Convolutional Neural Network followed by some chromosome post-processing. This step achieved 0.84 mean IoU, 0.9923 AP, and 100% individual chromosomes detection accuracy. The second stage is feature extraction and classification where we fine-tune VGG19 network using two different approaches, one by adding extra fully connected layer(s) and another by replacing fully connected layers with the global average pooling layer. The best accuracy obtained is 95.04%. The final step is detecting abnormality and this step obtained 96.67% average abnormality detection accuracy. To further validate the proposed classification method, we examined the Biomedical Imaging Laboratory dataset which is publicly available online and achieved 94.11% accuracy.