|
|
Journal of Zhejiang University SCIENCE A
ISSN 1673-565X(Print), 1862-1775(Online), Monthly
2026 Vol.27 No.3 P.288-305
Deep-learning-enabled automatic gene abnormality detection via fluorescence in situ hybridization
Abstract: Fluorescence in situ hybridization (FISH) is widely used for diagnosing cancer and genetic disorders due to its high specificity and accuracy. However, traditional methods face challenges such as suboptimal focus adjustments, subjective signal counting errors, and inefficiencies in imaging, limiting their use in high-throughput screening. To address these issues, we introduced the integrated FISH imaging and analysis system (FAST), an innovative solution that combines rapid filter switching, automated focusing, multilayer fluorescence signal fusion, and the improved ResNet152 deep learning framework. Compared with clinical manual counts and analysis of case reports of 10 patients with chronic lymphocytic leukemia (CLL), the FAST achieved an average nuclei segmentation accuracy of 98.28%. For abnormal gene detection, the model achieved an accuracy of 97.86%. Additionally, its intuitive interface allows the operator to complete the entire workflow—from scanning to report generation—within 45 min. FAST represents a significant advancement in cancer and genetic disorder diagnostics, offering a powerful tool for early detection.
Key words: Fluorescence in situ hybridization (FISH); Image enhancement; Nuclei segmentation; Fluorescence feature fusion; Abnormal gene classification
机构：1长春理工大学，计算机科学与技术学院，生命科学与技术学院，中国长春,130022；2吉林建筑大学，现代工业学院，中国长春，130119；3西安电子科技大学，杭州研究院，中国杭州,311231；4澳门大学，机电工程系，人工智能与机器人研究中心，中国澳门，999078；5不列颠哥伦比亚大学，电气与计算机工程系，加拿大温哥华，V6T 1Z4；6墨尔本大学，计算机与信息系统学院，澳大利亚帕克维尔，3010
目的：开发一种基于深度学习的基因异常检测自动化系统（FAST），以突破传统荧光原位杂交（FISH）技术在癌症和遗传疾病诊断中的局限性。
创新点：1.多层焦平面成像策略：增强不同焦深的成像，实现精确的细胞分析；2.图像增强与融合：结合多尺度滤波技术，提升信号清晰度；3.深度学习集成：优化ResNet152深度学习框架以进行高效的基因异常分类，细胞分割准确率为98.28%，基因异常检测准确率为97.86%。
方法：1.FISH成像：利用三色FISH探针和电动荧光显微镜进行多层成像；2.图像处理：基于多层荧光信号融合、图像增强以及细胞核和荧光信号的自动分割；3.AI模型：基于ResNet152的卷积神经网络（CNN）型，经过训练用于检测染色体异常（例如慢性淋巴细胞白血病中的13q14缺失）。
结论：1.FAST系统在细胞核分割和基因异常检测方面均取得高精度结果，细胞分割准确率达98.28%，以及基因异常检测准确率达97.86%。2.系统实现了从成像到报告生成的全流程自动化，且可在45分钟内完成检测，提升了诊断效率并降低了对人工经验的依赖，因此具有良好的临床应用潜力。
关键词组：
References:
Open peer comments: Debate/Discuss/Question/Opinion
<1>
DOI:
10.1631/jzus.A2500360
CLC number:
Download Full Text:
Downloaded:
1955
Download summary:
<Click Here>Downloaded:
9Clicked:
1776
Cited:
0
On-line Access:
2026-03-25
Received:
2025-07-29
Revision Accepted:
2025-11-01
Crosschecked:
2026-03-25