Combination of fluorescence sensor and artificial intelligence—a new method of quantitative ketamine detection

Abstract

Abstract Drug abuse is currently a growing concern, with the abuse of narcotic drugs, represented by ketamine, being particularly serious. Chronic use and overdose of such drugs can lead to hallucinations and serious health risks. Therefore, it is essential to establish a fast and high-precision detection method for strict control of drug abuse. In this study, we propose a fast detection process for small instruments, and the use of data augmentation and transfer learning techniques makes it possible to build a high-precision detection model based on small data sets. We also propose an EfficientNet-based Ketamine quantitative detection network. We maintain its lightweight architecture while introducing large kernel convolution to obtain a stronger feature representation without increasing the number of parameters. Our proposed detection process is time-saving and effective, reducing the time required to acquire large data sets, and our proposed model predicts all test data perfectly, solving the problem of bias of other networks in predicting low concentration samples, and breaking the limit of optical biosensor detection accuracy from 1 ng ml−1 to 0.1 ng ml−1.