[Paper Review] Image Segmentation to Distinguish Between Overlapping Human Chromosomes

Rihan Hu, Jeremy Karnowski+2 more|arXiv (Cornell University)|Dec 20, 2017

Genomic variations and chromosomal abnormalities17 references20 citations

TL;DR

This paper proposes a customized U-Net-based convolutional neural network for automated segmentation of overlapping human chromosomes in microscopic images. By leveraging pixel-wise classification with optimized architecture and hyperparameters, the model achieves 94.7% intersection over union (IoU) on overlapping regions and 88–94% on non-overlapping regions, significantly improving upon traditional manual and rule-based methods.

ABSTRACT

In medicine, visualizing chromosomes is important for medical diagnostics, drug development, and biomedical research. Unfortunately, chromosomes often overlap and it is necessary to identify and distinguish between the overlapping chromosomes. A segmentation solution that is fast and automated will enable scaling of cost effective medicine and biomedical research. We apply neural network-based image segmentation to the problem of distinguishing between partially overlapping DNA chromosomes. A convolutional neural network is customized for this problem. The results achieved intersection over union (IOU) scores of 94.7% for the overlapping region and 88-94% on the non-overlapping chromosome regions.

Motivation & Objective

• To automate the segmentation of overlapping human chromosomes in microscopic images, reducing reliance on manual or heuristic methods.
• To address the challenge of distinguishing between nearly identical, partially overlapping chromosomes in cytogenetic analysis.
• To develop a fast, accurate, and scalable deep learning solution for biomedical research and clinical diagnostics.
• To improve upon traditional methods such as thresholding, contour analysis, and manual cutting by leveraging neural networks.

Proposed method

• A customized U-Net architecture is used, with reduced depth (nearly one-third fewer layers) to optimize for small input images (88×88 pixels).
• The model is trained to predict pixel-wise segmentation labels: 0 (background), 1 (chromosome 1), 2 (chromosome 2), and 3 (overlapping region).
• Input images are preprocessed by correcting erroneous labels and cropping to 88×88 pixels to ensure divisibility by 2 for pooling layers.
• The network uses 'same' padding and avoids cropping within the network to preserve spatial resolution.
• Various hyperparameters were tested, including loss functions, activation functions, class weighting, and label encoding strategies (integer vs. one-hot).
• The model is trained on 64% of data, validated on 16%, and tested on 20% of the dataset of 13,000 grayscale chromosome images with ground-truth segmentation maps.

Experimental results

Research questions

• RQ1 Can a deep learning model achieve high-accuracy segmentation of overlapping human chromosomes without human intervention?
• RQ2 How does model performance vary across overlapping versus non-overlapping chromosome regions?
• RQ3 What architectural simplifications can maintain high performance while reducing computation for small, structured images like chromosome pairs?
• RQ4 How do different loss functions and label encoding strategies affect segmentation accuracy in this low-dimensional, high-similarity context?
• RQ5 Can the model be extended to handle more than two overlapping chromosomes or single chromosomes in future applications?

Key findings

• The model achieved an intersection over union (IoU) score of 94.7% on the overlapping region of chromosomes, indicating strong performance in distinguishing shared pixels.
• IoU scores for non-overlapping regions were 88.2% for one chromosome and 94.4% for the other, demonstrating high accuracy in segmenting individual chromosome parts.
• The model outperformed traditional methods such as histogram thresholding, geometric contour analysis, and manual cutting, which are error-prone and labor-intensive.
• The use of a simplified U-Net architecture reduced computational load while maintaining high accuracy, due to the small input size and limited shape variability of chromosomes.
• Data augmentation, including rotations, reflections, and stretching, is recommended to further improve generalization and model robustness.
• The model can be integrated into a pipeline with object detection to process full microscope images by first localizing chromosomes and then segmenting them.