M-AResNet: a novel multi-scale attention residual network for melting curve image classification

Pengxiang Su, Xuanjing Shen, Haipeng Chen, Di Gai, Yu Liu

Abstract: Melting curve image is a hallmark of quantitative polymerase chain reaction and is a crucial indicator for the validity of the cycle threshold. Current mainstream methods concentrate on analyzing the melting curve images via artificial process. Therefore, we design a novel multi-scale attention residual network, leveraging various levels space features for accurately classifying the melting curve images. Two modular components are designed in our algorithm. A multi-scale feature extraction module that consists of multi-parallel attention resnet units to selectively capture close related information from various scale feature maps while a series of adaptive multi-scale fusion modules to complete cross-subnet fusion of information. In addition, we also collect massive fluorescence signal data to draw melting curve images for constructing a novel dataset. Our method is evaluated on 3 different benchmark datasets including the self-constructed melting curve image dataset, heartbeat signal dataset and natural color image dataset, a significant highlight is that it achieves a 2.0% accuracy improvement over state-of-the-art in average.