Development of a rapid detection method for patch thickness based on machine vision and near infrared spectroscopy: a case study of curcumin patch.

Patches represent a significant dosage form within transdermal drug delivery systems. In the quality control of transdermal patches, thickness not only affects the drug loading per unit area and drug release behavior but also is associated with patient usability and product stability. Targeting curcumin patch, a method used for thickness distribution monitoring utilizing machine vision coupled with near infrared (NIR) spectroscopy was proposed. Twenty-nine batches of patches were prepared via the central composite design. The RGB color space data were extracted from the captured images of all patches. Subsequently, the color information and the NIR spectral data were fused as input dataset for model training. In this study, whale optimization algorithm (WOA), genetic algorithm, and competitive adaptive reweighted sampling were used for variable selection. To achieve the prediction of patch thickness, three modeling strategies, (a) partial least squares regression, (b) light gradient boosting machine (LightGBM) and (c) convolutional neural network with an attention mechanism, were adopted. The results indicated that the WOA-LightGBM model achieved commendable prediction performance with the relative prediction deviations value of 4.646 and a strong correlation coefficient of 0.954. The reliable prediction results and extremely short response time indicated the significant application potential for this model in real-time monitoring of patch thickness.