Application of Deep Convolutional Neural Networks and IR Spectroscopy for the Detection of Drugs and Toxins

This paper explores the use of deep learning architectures to identify and categorize infrared spectral data with the objective of classifying drugs and toxins with a high level of accuracy. The model proposed uses a custom convolutional neural network to learn the spectrum of 192 drugs and 207 toxins. Variations in the architecture and number of blocks were iterated to find the best possible fit. A real-time implementation of such a model faces a lot of issues such as noise from different sources, spectral magnitude off-setting, and wavelength rotation. This paper aims to tackle some of these problems. Another common issue is the use of extensive pre-processing which makes it difficult to automate the entire process. We have aimed to side-step this issue with the architecture proposed. The focus is on 2 applications - detection of drugs and toxins. The data sets used are from different sources, each with its own noise factor and sampling rate. Some of the traditional models like Principal Component Analysis (PCA) and Support Vector Machines (SVM) were also tested on the datasets. The model works with minimal input data of two spectra (and three augmentations of the same) to learn the features and classifies the data from a source independent of the input. The proposed model showed a significant improvement in accuracy when compared to the other models currently in use, achieving an overall accuracy of 96.55\%. The model proposed performs extremely well with a minimal sampling rate and shows no loss in accuracy of classification even with an increase in the number of classes. The research conducted has the scope of being extended to the identification of counterfeit drugs which is a growing cause for concern. Another application could be in the detection of the presence of harmful toxins.