A scalable deep attention mechanism of instance segmentation for the investigation of chromosome.

Journal: SLAS technology
Published Date: May 11, 2025

Abstract

Chromosome segmentation in metaphase images is a critical yet challenging task in cytogenetics and genomics due to the inherent complexity, variability in chromosome shapes, and the scarcity of high-quality annotated datasets. This study proposes a robust instance segmentation framework that integrates an automated annotation pipeline with an enhanced deep learning architecture to address these challenges. A novel dataset is introduced, comprising metaphase images and corresponding karyograms, annotated with precise instance segmentation information across 24 chromosome classes in COCO format. To overcome the labor-intensive manual annotation process, a feature-based image registration technique leveraging SIFT and homography is employed, enabling the accurate mapping of chromosomes from karyograms to metaphase images and significantly improving annotation quality and segmentation performance. The proposed framework includes a custom Mask R-CNN model enhanced with an Attention-based Feature Pyramid Network (AttFPN), spatial attention mechanisms, and a LastLevelMaxPool block for superior multi-scale feature extraction and focused attention on critical regions of the image. Experimental evaluations demonstrate the model's efficacy, achieving a mean average precision (mAP) of 0.579 at IoU = 0.50:0.95, surpassing the baseline Mask R-CNN and Mask R-CNN with AttFPN by 3.94% and 5.97% improvements in mAP and AP50, respectively. Notably, the proposed architecture excels in segmenting small and medium-sized chromosomes, addressing key limitations of existing methods. This research not only introduces a state-of-the-art segmentation framework but also provides a benchmark dataset, setting a new standard for chromosome instance segmentation in biomedical imaging. The integration of automated dataset creation with advanced model design offers a scalable and transferable solution, paving the way for tackling similar challenges in other domains of biomedical and cytogenetic imaging.

Authors

  • Neelam Umbreen
    National University of Sciences and Technology (NUST), Islamabad, Pakistan. Electronic address: neelamumbreen.phd@smme.edu.pk.
  • Sara Ali
    Department of Nutrition & Dietetics, School of Health Sciences, University of Management and Technology, Lahore, Pakistan.
  • Hasan Sajid
    National University of Sciences and Technology (NUST), Islamabad, Pakistan. Electronic address: hasan.sajid@smme.nust.edu.pk.
  • Yasar Ayaz
    National Center of Artificial Intelligence (NCAI), National University of Sciences and Technology (NUST), Islamabad, Pakistan.
  • Shrooq Alsenan
    Research Center, College of Computer and Information Sciences, Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia; Research Chair in Healthcare Innovation, Information Systems Department, College of Computer and Information Sciences, King Saud University, Riyadh, Saudi Arabia. Electronic address: 436203869@student.ksu.edu.sa.
  • Yunyoung Nam
  • So Yeon Kim
  • Muhammad Baber Sial
    School of Automation Science and Electrical Engineering, Beihang University, Beijing 100191, China. Electronic address: babersial@buaa.edu.cn.

Keywords

No keywords available for this article.

External Resources

Popular Topics
Recent Journals