Feature Fusion Classifier With Dynamic Weights for Abnormality Detection of Amniotic Fluid Cell Chromosome

Abstract

Chromosomal karyotype is important to determine whether a newborn has a genetic disorder. There are two main categories of chromosomal abnormalities: structural abnormalities, in which the chromosome structure is altered, and chromosome number abnormalities. Manual karyotyping is complex and takes a lot of time because it requires a high degree of domain expertise. Based on this investigation, we propose a new method of chromosome defect detection based on deep learning with 20,299 chromosome images from Dongguan Kanghua Hospital as data that integrates the diversity of chromosome features and trains a classifier model based on feature fusion for chromosome abnormality detection. We put forward a feature fusion classifier with dynamic weights (FFCDW) for chromosomal abnormality detection, after data augmentation with three deep learning networks, ResNet, SENet, and VGG19, the three trained models are combined using a dynamic weighting approach. Experiments prove the FFCDW method outperforms these mainstream models of ResNet, SENet, and VGG19. The proposed method based on FFCDW achieves a precision of 0.8902 and an F1-score of 0.8805 with a small standard deviation (0.00903 and 0.00892, respectively). In addition, the algorithm can automatically assign weights based on the results of a single model, and the strategy with dynamic weights outperforms the strategy with fixed weights in the proposed feature fusion classifier.

Keywords

• Chromosome karyotype analysis
• deep learning
• machine vision
• dynamic weighting
• feature fusion classifier
• model fusion