Vision Transformer on SERS maps

This repository provides the implementation for our paper Nitroaromatic explosives detection and quantification using attention-based transformer on surface-enhanced Raman spectroscopy maps. We experimentally demonstrated that we significantly outperform or are on par with the existing approaches for explosives detection and quantification using raw SERS maps as input. We also open-source three SERS datasets measured down to very low concentrations: 4-NBT, 4-dinitrophenyl hydrazine (DNP), and Picric acid (PA) at https://data.dtu.dk/articles/dataset/Datasets_for_replicating_paper_Nitroaromatic_Explosives_detection_and_quantification_using_Attention-based_Transformer_on_surface-enhanced_Raman_spectroscopy_maps_/21602928.

Requirement

git clone https://github.com/lyn1874/molecule_detection_quantification.vit
cd molecule_detection_quantification
conda env create -f molecule_env.yml 
conda activate torch_dl

Data preparation

Download the dataset into folder datasets/

Model preparation

Download the model checkpoints into exp_data/VIT/

Model testing

Please look at the file test_experiment.ipynb for evaluating the experiment

Model training

ViT-based explosives detection and quantification

./run_multiple_vit.sh detection quantification dataset gpu_index 
Args:
	detection: bool variable, true/false 
	quantification: bool variable, true/false
	dataset: str variable, TOMAS/DNP/PA
	gpu_index: int, which gpu to use

For example, to run a detection experiment on dataset DNP, you can simply run: ./run_multiple_vit.sh true false DNP 0

Spectra-based explosive detection and quantification

./run_multiple_spectra.sh dataset model_group detection quantification version percentage gpu_index 
Args:
	dataset: str varialbe, TOMAS/DNP/PA 
	model_group: str variable, xception/unified_cnn/resnet 
	detection: bool variable, true/false 
	quantification: bool variable, true/false 
	version: int 
	percentage: float, 0.002/0.005/0.01/0.02/0.05/0.1/0.2/0.5/1.0 
	gpu_index: int 

For example, to run a detection experiment on dataset PA with model xception, and you choose to average each SERS map with the Top-2% of the spectra that have the highest peak intensity: ./run_multiple_spectra.sh PA xception true false 0 0.002 0

SERS map generation

The processes for generating SERS maps are shown in the jupyter file SERS_maps_generation.ipynb

Citation

If you use this code, please cite:

@Article{D3AN00446E,
author ="Li, Bo and Zappalá, Giulia and Dumont, Elodie and Boisen, Anja and Rindzevicius, Tomas and Schmidt, Mikkel N. and Alstrøm, Tommy S.",
title  ="Nitroaromatic explosives’ detection and quantification using an attention-based transformer on surface-enhanced Raman spectroscopy maps",
journal  ="Analyst",
year  ="2023",
publisher  ="The Royal Society of Chemistry",
url  ="http://dx.doi.org/10.1039/D3AN00446E",
}