A two-stage hybrid framework for predicting fine particle fraction from SEM images using deep feature extraction and class-specific regression.

Fine particle fraction (FPF), a key predictor of inhalation powder efficacy, is conventionally measured using expensive, time-consuming in vitro methods that require specialized equipment. This study presents a novel approach for predicting FPF from scanning electron microscopy (SEM) images using deep learning and machine vision. The dataset comprised 764 preprocessed SEM images and their corresponding FPF values. A comparison of nine pretrained architectures identified DenseNet169 as the best feature extractor. A novel two-stage pipeline was then developed. In the first stage, the fine-tuned DenseNet169 model classified the images into four FPF classes. In the second stage, a random forest algorithm performed class-specific regression using the extracted feature vectors. The developed model predicted FPF with a weighted mean absolute error (MAE) of 1.84, representing an approximately 90% reduction in error compared with direct regression models, which produced errors of approximately 20. Statistical analysis indicated that the test set adequately represented the full dataset, and the model performed well on it. Grad-CAM heatmaps showed that the model focused on relevant regions of particle morphology. The proposed approach may facilitate rapid, cost-effective screening of inhalation formulations. By reducing material consumption, energy use, and laboratory waste, this method may also support industrial quality-control workflows and the development of dry powder inhalers.