Segmentation of Overlapping Chromosomes Using Kriging based U-Net

A significant challenge in cytogenetics is the identification of overlapping chromosome images owing to the complex shapes and occlusions present in these images. This work introduces a novel approach that integrates Kriging-Based attention mechanisms with the U-Net architecture to improve the segmentation accuracy of overlapping chromosomes in metaphase images. The attention mechanism based on kriging exploits spatial correlation information to adaptively emphasize regions with high relevance. The proposed method employs the encoder-decoder structure of U-Net, with the Kriging attention mechanism applied to the feature maps in the bottleneck layer, refining spatial dependencies and improving feature representation. A comprehensive set of experiments is conducted using publicly available chromosome metaphase image datasets to evaluate the effectiveness of the Kriging attention module, comparing it with traditional U-Net and attention-based models.The proposed method could successfully segment various types of occlusions present in metaphase images.Results show a significant improvement in segmentation performance, with better Dice coefficient and Intersection over Union (IoU) scores, demonstrating that Kriging attention enhances the ability to handle overlapping chromosome structures while preserving fine structural details. This approach opens the possibilities for more accurate automated analysis of chromosome images, facilitating applications in cytogenetics and clinical diagnostics.