Drug classification with a spectral barcode obtained with a smartphone Raman spectrometer

Miniaturizing optical spectrometers for portable applications is a significant area of research. Smartphones, with their ubiquitous nature and advanced processing capabilities, offer an ideal platform for integrating such devices. Existing smartphone-based spectrometers often utilize gratings, which are challenging to miniaturize. This research addresses this challenge by developing a compact Raman spectrometer using a CMOS image sensor and a periodic array of bandpass filters to directly capture a 2D Raman spectral intensity map, termed a "spectral barcode." This approach simplifies spectral acquisition and analysis, making it particularly suitable for applications such as drug identification, where counterfeit drugs pose a substantial threat to public health. Current methods, relying on drug name, shape, color, or RGB image comparisons, offer insufficient accuracy. Raman spectroscopy, providing molecular fingerprints, offers a more reliable approach. While some research has explored Raman spectroscopy for drug classification using machine learning, limitations exist in terms of spectral resolution and miniaturization. This study overcomes these limitations by integrating a miniaturized Raman module with a smartphone camera, creating a portable and user-friendly system for rapid and accurate drug classification.

Numerous studies have explored the use of smartphones as platforms for various optical spectroscopy techniques. These include applications in food inspection, beauty care, and healthcare, leveraging the smartphone's camera and processor for on-device or cloud-based analysis. While many smartphone-based spectrometers utilize gratings for spectral dispersion, the inherent difficulties in miniaturizing gratings have led researchers to explore alternative approaches such as photonic crystals, metasurfaces, quantum dots, and silicon nanowires integrated onto image sensors. However, these approaches often compromise spectral resolution or require complex numerical analysis to recover the input spectrum. The challenge of counterfeit drug identification has also been addressed by existing smartphone apps, which primarily rely on visual features (name, shape, color) or RGB image comparisons against databases. These methods have limited accuracy due to visual similarities or database limitations. Previous research in drug classification using Raman spectroscopy and machine learning techniques like PLS-DA, PCA, and CNNs has shown promising results, but often lacked the portability and ease of use offered by smartphone integration.

The researchers developed a smartphone-based Raman spectrometer by attaching a compact external module containing a 785 nm laser diode to the rear camera of a Samsung Galaxy Note 9. The module excites Raman signals from the sample, which are then captured by the smartphone's CMOS image sensor equipped with a 2D periodic array of 120 bandpass filters, covering the range of 830–910 nm. The captured 2D Raman intensity map is defined as a "spectral barcode." The fabrication of the bandpass filters involved plasma-enhanced chemical vapor deposition and photolithography, creating a pair of Si/SiO2 distributed Bragg reflectors separated by a Si cavity layer to adjust the resonant wavelength. The spectral barcode contains approximately 1200 bits of information. The data processing involved converting the raw image into a normalized spectral barcode, analogous to conventional barcodes. A convolutional neural network (CNN) with a simplified ResNet architecture was developed and trained to classify the spectral barcodes of 54 common drugs for various diseases and over-the-counter medications, based on their major chemical components. The CNN architecture included a residual block with convolution, batch normalization, addition, and ReLU activation, followed by two fully connected layers. For brand name identification, another CNN with a similar architecture, but a different fully connected layer size, was employed. The RGB images of the drugs, captured by the smartphone camera, were also used as additional input to a separate CNN with a VGGNet architecture to classify shapes and colors, improving the accuracy of brand identification. The combined output of both CNNs (spectral barcode and RGB) was used to predict the brand name, with the outputs treated equally.

The smartphone Raman spectrometer achieved high accuracy in drug classification. The CNN trained on the spectral barcodes achieved 99.0% accuracy in classifying the major chemical components of 54 drugs. The model could also identify the components of four additional drugs not included in the training dataset (99.8% accuracy). Brand name identification based on spectral barcodes alone reached 79.5% accuracy, while incorporating RGB image information improved the accuracy to 83.2%. The study demonstrated that even drugs with similar appearances could be distinguished based on their unique spectral barcodes, providing a valuable tool for combating counterfeit drugs.

This research successfully demonstrates a novel approach to drug classification using a miniaturized, smartphone-integrated Raman spectrometer. The use of spectral barcodes simplifies spectral data acquisition and analysis, making it highly suitable for real-world applications, particularly for identifying counterfeit medications. The high classification accuracy achieved for both drug components and brand names validates the effectiveness of the approach. The integration of RGB image information further enhances the accuracy of brand identification, showcasing the potential for multi-modal data fusion to improve overall performance. The ability to classify drugs not included in the training dataset highlights the generalizability and robustness of the developed CNN model. This technology has broad implications for healthcare, pharmaceutical quality control, and law enforcement, enabling rapid and accurate drug identification in various settings.

This study introduced a novel approach for drug classification using a smartphone-integrated Raman spectrometer and spectral barcodes. The high accuracy achieved in identifying both drug components and brand names demonstrates the potential of this technology for combating counterfeit drugs and improving healthcare practices. Future work could focus on increasing the spectral range and resolution, exploring advanced spectral processing techniques to enhance accuracy, and incorporating additional features such as imprinted drug markings for enhanced identification. Expanding the database to include a more comprehensive range of drugs would further improve the generalizability of the system.

The current system relies on a specific smartphone model and requires careful positioning of the drug sample for optimal Raman signal acquisition. The accuracy of brand name identification, while improved with RGB image fusion, could be further enhanced by incorporating additional information about drug coatings or additives. The relatively small dataset used for training and testing might limit the generalization of the results to other drug samples and populations. Further research is needed to address these limitations and enhance the robustness and applicability of the system.