Publication: Gfp pattern recognition in raman spectra by modified vgg networks for localisation tracking in living cells
Issued Date
2021-01-01
Resource Type
ISSN
01258281
Other identifier(s)
2-s2.0-85102677126
Rights
Mahidol University
Rights Holder(s)
SCOPUS
Bibliographic Citation
Engineering Journal. Vol.25, No.2 (2021), 151-160
Suggested Citation
Nungnit Wattanavichean, Jirasin Boonchai, Sasithon Yodthong, Chakkrit Preuksakarn, Scott C.H. Huang, Thattapon Surasak Gfp pattern recognition in raman spectra by modified vgg networks for localisation tracking in living cells. Engineering Journal. Vol.25, No.2 (2021), 151-160. doi:10.4186/ej.2021.25.2.151 Retrieved from: https://repository.li.mahidol.ac.th/handle/20.500.14594/76989
Research Projects
Organizational Units
Authors
Journal Issue
Thesis
Title
Gfp pattern recognition in raman spectra by modified vgg networks for localisation tracking in living cells
Abstract
The coupling between Raman spectroscopy and green fluorescent protein (GFP) labelling informs chemical compositions at the specific sites. This information leading to study that explain core knowledge of living organism and eventually advance our conventional technique of medical diagnosis. In order to achieve these purposes, the precise interpretation is required. A massive number of Raman/GFP spectra as well as identification of GFP contribution in each spectrum are arroaches to achieve those goals. In the paper, CNN is proposed to classify the spectra with and without GFP signal. The dataset of GFP-positive and GFP-negative spectra were created with various size and background color. The feature extraction and classification are conduced with VGG networks. To increase the performance of VGG network, the modified VGG13 and modified VGG19 were designed. These two models extend fully-connected layer from 3 (the original VGG model) to 5 layer for better classification task. Batch normalization is also added at the end of feature extraction units to reduce unpredicted shifting of parameters. The original VGG16, VGG19, and ResNet50 are used as comparison models. The results show that both of our modified VGG models significantly enhances training accuracy of the network comparing to the original VGG. The accuracy of original VGG can be increased when applied pre-trained weight, but the accuracies are yet slightly lower than modified models. Training on ResNet, deeper network, gave the comparable accuracy with our modified models.