Down Syndrome detection with Swin Transformer architecture

Objective: Down Syndrome, also known as Trisomy 21, is a severe genetic disease caused by an extra chromosome 21. For the detection of Trisomy 21, despite those statistical methods have been widely used for screening, karyotyping remains the gold standard and the first level of testing for diagnosis. Due to karyotyping being a time-consuming and labour-intensive procedure, Computer Vision methodologies have been explored to automate the karyotyping process for decades. However, few studies have focused on Down Syndrome detection with the Transformer technique. This study develops a Down-Syndrome-Detector (DSD) architecture based on the Transformer structure, which includes a segmentation module, an alignment module, a classification module, and a Down Syndrome indicator. Methods: The segmentation and classification modules are designed by homogeneous transfer learning at the model level. Transfer learning techniques enable a network to share weights learned from the source domain (e.g., millions of data in ImageNet) and optimize the weights with limited labeled data in the target domain (e.g., less than 6,000 images in BioImLab). The Align-Module is designed to process the segmentation output to fit the classification dataset, and the Down Syndrome Indicator identifies a Down Syndrome case from the classification output. Results: Experiments are first performed on two public datasets BioImLab (119 cases) and Advanced Digital Imaging Research (ADIR, 180 cases). Our performance metrics indicate the good ability of segmentation and classification modules of DSD. Then, the DS detection performance of DSD is evaluated on a private dataset consisting of 1084 cells (including 20 DS cells from 2 singleton cases): 90.0% and 86.1% for cell-level TPR and TNR; 100% and 96.08% for case-level TPR and TNR, respectively. Conclusion: This study develops a pipeline based on the modern Transformer architecture for the detection of Down Syndrome from original metaphase micrographs. Both segmentation and classification models developed in this study are assessed using public datasets with commonly used metrics, and both achieved good results. The DSDproposed in this study reported satisfactory singleton case-specific DS detection results. Significance: As verified by a medical specialist, the developed method may improve Down Syndrome detection efficiency by saving human labor and improving clinical practice.