染色体的分类是核型分析的重要任务之一.因其柔软易弯曲,且类间差异小、类内差异大等特点,其精准分类仍然是一个具有挑战性的难题.对此,提出一种基于网格重构学习(Grid reconstruction learning,GRiCoL)的染色体分类模型.该模型首先将染色体图像网格化,提取局部分类特征;然后通过重构网络对全局特征进行二次提取;最后完成分类.相比于现有几种先进方法,GRiCoL同时兼顾局部和全局特征提取更有效的分类特征,有效改善染色体弯曲导致的分类性能下降,参数规模合理.通过基于G带、荧光原位杂交(Fluorescence in situ hybridization,FISH)、Q带染色体公开数据集的实验表明:GRiCoL能够更好地弱化染色体弯曲带来的影响,在不同数据集上的分类准确度均优于现有分类方法.
Abstract
Chromosome classification is one of the key tasks of karyotype analysis.However,due to chromosomes are flexible hence exhibit less difference between different types while significant difference within same type,accur-ate classification of chromosome remains a challenging issue.In this paper,a chromosome classification model based on grid reconstruction learning(GRiCoL)is proposed.To weaken the impact of the bendy state,chromosome im-ages are first grid-enabled for feature extraction separately.Subsequentially,global features are extracted for the second time by reconstruction network,which is followed by classification.Compared with the state-of-the-art methods,the proposed GRiCoL can get more efficient discriminable features based on both local and global fea-tures,therefore can overcome the adverse effects of bandy form of chromosome with reasonable parameter scale.Ex-periments on public G band,fluorescence in situ hybridization(FISH)as well as Q band chromosome datasets show that GRiCoL can extract discriminative features that weaken the bending of chromosomes more efficiently,mean-while,higher performance was obtained as compared to current classification algorithms.
维普
万方数据