Development of computer aided Classification and centromere Detection system for g band human chromosome images

Abstract

The research work focuses on developing automatic classification and centromere detection from G-band human chromosome images to recognize common abnormalities. The proposed research comprises three main processes: segmentation, feature extraction, and classification. The objective of chromosome segmentation is to remove irrelevant objects and separate single chromosomes from chromosome clusters. Separation of touching and overlapping chromosomes, as well as background noise and irrelevant objects, are observed to be challenges. Hence, the fully-automated raw G-band chromosome image segmentation is proposed to resolve the above-said difficulties, with the objective of segmenting every single chromosome. The background noise and extraneous objects are initially eliminated using their characteristics, and all chromosomes and clusters are separated from the background using Connected Component Analysis (CCA). In the classification stage, Deep learning classifiers are the best approach for automatic karyotyping and additional abnormality research. VGG (Visual Geometry Group) is one of the most widely used image recognition architectures. To improve model stability and reduce training time, the VGG-16 network model is used for experimentation. It has 16 layers, including 13 convolution layers, a 3 x 3 convolution filter, and three fully connected layers (FC). The VGG-16 network is enhanced by integrating the Global Average Pooling (GAP) and Fully Connected (FC) layers, as well as the Batch Normalization (BN) layer for efficient feature extraction and chromosomal classification. The research focuses on automated segmentation, classification, identification of centromeres, and numerical abnormalities. The future objective of the research project is to build an automated karyotype for chromosome classification to detect structural chromosome abnormalities. newline