(587f) The Development of a High-Precision Inline Image Analysis Method for Acquiring Emulsion Dynamics

Emulsions are widely applied in industries such as food, pharmaceuticals, cosmetics, and paint. Understanding the droplet size is crucial for controlling and optimizing emulsion processes. Methods for obtaining the droplet size from the image processing of microscopic images obtained via inline measurements are known to overcome existing challenges (such as that posed by excluding the effects of dilution by laser diffraction or spectroscopy), and they require less time for measurement. However, it remains difficult to accurately segment and extract individual droplets in the cases of high-concentration emulsions, due to droplet overlap and agglomeration. In this study, we focused on image processing based on deep learning [1, 2]. With the use of deep learning, it is possible to accurately segment individual droplets and obtain the droplet-size distribution. A challenge arises when the number of data is insufficient, making it difficult to properly segment individual droplets. In this study, we developed a high-precision inline image analysis method for acquiring emulsion dynamics that can identify high-precision droplet-size distributions with the use of only 10 microscopic images. We also describe our evaluation of the new method's accuracy and the dynamics of droplet splitting and coalescence behavior under various mixing conditions.

Methods

An oil/water (O/W) emulsion with a water-to-oil ratio of 7:3 was used. An inline microscope was inserted into a 1-L lab-scale mixing tank, and mixing was performed. Microscopic images were obtained at 2-sec intervals while the mixing was stopped (static state). We randomly selected 10 images from among the microscopic images with various droplet-size distributions obtained under different mixing conditions. After identifying the regions of bubbles and droplets in the images, we made annotations to specify the regions of each dispersed phase in the images at the pixel level. For the training data, we constructed a model that was trained using a type of deep learning, i.e., instance segmentation, to predict individual droplets on a pixel-by-pixel basis.

Verification of the constructed image-processing algorithm

We compared the accuracy of directly training an instance segmentation model on annotated data of bubble and droplet regions with the results obtained by applying the image-processing algorithm constructed in this study. The constructed image-processing algorithm builds instance segmentation models for bubbles and droplets in the original image, as well as for droplets in the segmented images obtained by dividing the original image. By combining the output results of the three models, the algorithm detects only the droplets. As a result, it was possible to solve two issues: the misdetection of bubbles as droplets that occurs when directly training the instance segmentation model, and the decrease in detection accuracy for small droplets.

For a comparison of the new method's accuracy with those of existing measurement methods, we compared the results obtained with a laser diffraction particle size analyzer (LD) with the results of the newly constructed method. An inline microscope was inserted into a 30-L bench stirrer tank, and an O/W emulsion with a unimodal droplet-size distribution was prepared by high mixing for 10 min. Our comparison of the measurement results obtained with the constructed technique with those of the laser diffraction particle size analyzer for the obtained emulsion confirmed that the measurement results of the two methods were generally consistent.

The acquisition of droplet breakup and coalescence dynamics using the new method

To acquire the dynamics of droplet breakup and coalescence, we inserted an inline microscope into a 30-L bench-scale mixing tank, and we observed the temporal changes in the non-steady-state droplet-size distribution by taking images at 2-sec intervals. For the capture of the coalescence behavior, the changes in droplet size were tracked during a 5-min rest period after 10 min of mixing. For the examination of the breakup behavior, the changes in droplet size were tracked during the 150-min period of mixing from the start.

When we investigated the change in the droplet-size distribution during the static state under different mixing conditions, we observed that in systems with broad droplet-size distributions under low mixing conditions, the tendency for the standard deviation of the droplet-size distribution to increase over time was successfully captured. When we tracked the change in droplet size during mixing, we set and compared two mixing temperature conditions. In both cases, the droplet size decreased with mixing and converged after a certain period. The comparison of the statistical quantities of the droplet-size distribution under different mixing temperatures confirmed that the difference in mixing temperature was prominently reflected in the larger-droplet side of the droplet-size distribution, i.e., >90% diameter.

Conclusion

Our new method successfully obtained droplet-size distributions with an accuracy comparable to that of a laser diffraction particle size analyzer using only 10 images. The method also successfully captured the differences in droplet size in droplet coalescence during the static state, the droplet breakage during mixing, and these values under different mixing conditions (mixing speed and temperature). A high-precision method for acquiring emulsion dynamics focusing on droplet size changes has thus been successfully constructed. In the future, by applying this technology to emulsification and other fields, we can expect to obtain more detailed dynamics through image analyses.

References

[1] H. Khosravi et al., Int. J. Pharm., 649,123633 (2024).

[2] J. Liu et al., Chem. Eng. J., 451, 4 (2023).