Karyotyping is the most commonly used genetic tool for diagnosing diseases associated with chromosomal abnormalities. It generates images of the chromosomes of a patient in which quantity or shape discrepancies against normal chromosomes might suggest chromosomal abnormalities. However, the current methods are cumbersome and require manual or half-automatic separation of overlapping chromosomes, significantly limiting the productivity of clinical geneticists and cytologists. In this project, we implemented a fully automatic method, called ChromSeg, which efficiently separates crossing-overlap chromosomes. It uses a new neural network architecture called “region-guided UNet++” to accurately detect crossing-overlap chromosomes from metaphase cell images. A new heuristic algorithm, called “crossing-partition”, is then applied to splice and reconstruct the crossing-overlap chromosomes into single chromosomes. While there are a very limited number of publicly accessible annotations on overlapping chromosomes, we manually annotated 345 images for our model training and performance testing. Benchmarking results showed that our method achieved 99.1% overlap detection on crossing-overlap chromosomes and outperformed the second best method by 3.1%. Notably, this is the first tool to provide an image of the reconstructed chromosomes; other tools provide only segmentation suggestions, which are of less value to end-users. The source code of ChromSeg is available at https://github.com/HKU-BAL/ChromSeg, and the 345 annotated images are available at http://www.bio8.cs.hku.hk/bibm/.
Stage 1 use region-guided UNet++ to localize and segment crossing-overlap regions for chromosomes clusters.
Stage 2 use crossing-partition algorithm to splice and reconstruct non-overlap parts and crossing-overlap parts.