ARMS Net: Overlapping chromosome segmentation based on Adaptive Receptive field Multi-Scale network. (July 2021)

• Record Type: Journal Article
• Authors: Wang, Guangjie; Liu, Hui; Yi, Xianpeng; Zhou, Jinjun; Zhang, Lin
• Is Part Of: Biomedical signal processing and control. Volume 68(2021)
• Publication Date: 2021-07
• Subjects: Karyotype analysis -- Chromosome segmentation -- Adaptive receptive field -- Multi-scale; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.102811
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: An adaptive receptive field multi-scale network (ARMS Net) based on UNet architecture is proposed to segment chromosomes that curled and overlapped together randomly in different images. The number of pooling operations is optimized to balance the requirements of deep semantic information extraction and high precision segmentation. The adaptive multi-scale feature extraction module is designed to replace the standard convolution at the bottom of UNet, such that the receptive fields can adaptively match the size of feature map. An adaptive smooth weighted cross entropy loss function is defined to resolve sample category unbalance issue. Abstract: Karyotype analysis has become the crux for diagnosis of genetic diseases, which requires automatic precise identification of quantitative and structural abnormalities. Thus, segmentation of chromosomes from microscope captured photos by image processing technique can help promote the automatic identification of chromosomes. However, chromosomes are often curled and overlapped together randomly in different images, which makes overlapping chromosome segmentation the hot topic in karyotype analysis. Herein an Adaptive Receptive field Multi-Scale network (ARMS Net) based on UNet architecture is proposed. The number of pooling operations is optimized to balance the requirements of deep semantic information extraction and high precision segmentation. The adaptive multi-scale feature extraction module is designed to replace the standard convolution at the bottom of UNet, such that the receptive fields can adaptively match the size of feature map. Besides, an adaptive smooth weighted cross entropy loss function is defined to resolve category imbalance issue. Experimental results show that the Intersection of Union (IoU) score of ARMS Net segmented overlapping area is 99.45%, which is 3.2% higher than that achieved by UNet (96.38%), and 10.1% higher than that achieved by CE-Net (90.35%). In a word, ARMS Net is expected to be used as the backbone network for chromosome instance segmentation in its end-to-end identification.