(T1030-12-68) Applications of Deep Learning to Predict Aerosol Performance of Dry Powders for Inhalation

Purpose: Dry powder inhalers (DPI) are widely used devices to treat respiratory diseases. Aerosol performance of dry powders for inhalation is important for the product development process and is typically characterized by fine particle fraction (FPF) and mass median aerodynamic diameter (MMAD). However, a conventional trial-and-error method for formulation screening and optimization is time-consuming and laborious. Therefore, we developed a deep learning-based method to estimate the aerosol performance based on the properties of the powder made by thin-film freezing, specifically using image analysis.

Methods: In this study, a collection of 132 scanning electron microscopy (SEM) images was initially obtained from published literature and additional experiments ^1–3. The SEM images were manually classified into acceptable formulations (FPF of recovered dose ≥ 60%) and unacceptable formulations (FPF of recovered dose < 60%). Several image processing techniques including resizing, adaptive histogram equalization algorithm (CLAHE), and augmentation, were performed prior to the model build-up. A convolutional neural network (CNN) implemented through TensorFlow was then applied for the SEM image classification ⁴. In addition, 316 SEM images of dry powders prepared by thin-thin freezing for three different drugs and their formulations including tacrolimus, voriconazole, and remdesivir, were used to validate the CNN model. Several metrics including accuracy, receiver operating characteristic (ROC), and area-under-the-curve (AUC) were used to evaluate the model.

Results: We found that applying CLAHE to SEM images is an effective pre-processing step. It not only improved the image contrast but also enhanced the CNN model performance from an accuracy of 0.677 to 0.95 in validation subset. After the image augmentation, a balanced image data set containing 4266 SEM images was obtained and used for the training of the CNN model. The CNN model performed well in the testing subset with an accuracy of 0.839 and AUC of 0.927. Furthermore, the CNN model exhibited satisfactory performance with accuracies of 0.948, 0.730, and 0.821 for tacrolimus, remdesivir, and voriconazole formulations, respectively.

Conclusion: In this study, we have successfully developed a deep-learning-based aerosol performance analyzing method and achieved high accuracy and adaptability. Most importantly, this study can be used to reduce the workload of conventional experimental analysis. In addition, we plan to continue this study and investigate few-shot learning to further reduce the amount of data requirement. In addition, we plan to continue this study and investigate few-shot learning to further reduce the amount of data requirement.

References: 1. Sahakijpijarn S, Moon C, Ma X, et al. Using thin film freezing to minimize excipients in inhalable tacrolimus dry powder formulations. International Journal of Pharmaceutics. 2020;586:119490. doi:10.1016/J.IJPHARM.2020.119490
2. Moon C, Sahakijpijarn S, … JKJ of DD, 2019 undefined. Processing design space is critical for voriconazole nanoaggregates for dry powder inhalation produced by thin film freezing. Elsevier. Accessed March 18, 2022. https://www.sciencedirect.com/science/article/pii/S1773224719308263?casa_token=7_Qwkl7wbzAAAAAA:3DxXjdx_dmwr4tIMQc9T6gT-KdwNYiyREgiccMZ1N47GYjV3PBeKc576dpBU_tK6F4WKoCcWB6w
3. Sahakijpijarn S, Moon C, Koleng JJ, Christensen DJ, Williams RO. Development of Remdesivir as a Dry Powder for Inhalation by Thin Film Freezing. Pharmaceutics. 2020;12(11):1-27. doi:10.3390/PHARMACEUTICS12111002
4. TensorFlow. Accessed January 27, 2022. https://www.tensorflow.org/

Acknowledgements: This study was partially supported by a sponsored research agreement with TFF Pharmaceuticals, Inc. The University of Texas System has licensed IP on thin-film freezing to TFF Pharmaceuticals, Inc. Williams reports financial support from TFF Pharmaceuticals. Williams reports a relationship with TFF Pharmaceuticals, Inc. that includes consulting or advisory, equity or stocks, and research funding. Moon reports a relationship with TFF Pharmaceuticals, Inc. that includes consulting or advisory.


Figure 1. (A)The training and validation process of the CNN model. CNN accuracy results for images processed by CLAHE. (B) CNN loss results for images processed by CLAHE. (C) CNN accuracy results for images without CLAHE processing. (D) CNN loss results for images without CLAHE processing.


Figure 2. Performance of CNN model. Confusion matrix (Left) and receiving operating characteristic (ROC) (Right)


Table 1. Accuracy summary of different TFF formulations