Deep Learning-Based Machine Vision System for Use in High-Shear Granulation

Aim

Develop a prototype machine vision system and evaluate feasibility of a deep learning-based image analysis approach for estimation of the state of granulation on the acquired granule bulk surface images.

Deep Learning-Based Machine Vision System for Use in High-Shear Granulation_aim

Introduction

Despite many decades of research in granulation technology, the high-shear granulation process is still poorly understood. Practical knowledge of technologists suggests that granule appearance largely indicates the process state. As an alternative to subjective visual examination, machine vision could provide a more reliable and continuous solution, which was already demonstrated on the lab scale [1].

Figure 1: Granule appearance at various states of granulation.

Materials and methods

Granules

Commercially produced via high-shear granulation on production-scale.
Bulk composition – 55 %w/w API (BCS class II), lactose monohydrate, low substituted hydroxypropyl cellulose, hydroxypropyl cellulose, sodium lauryl sulfate.
Granulation liquid – purified water.
Sifted in order to obtain 12 size classes: 40 µm, 102.5 µm, 152.5 µm, 215.0 µm, 302.5 µm, 377.5 µm, 450.0 µm, 550.0 µm, 655.0 µm, 780.0 µm, 925.0 µm, 1200.0 µm.

Imaging setup

Figure 2. The prototype machine vision system.

The prototype machine vision system is designed for installation on granulator sight glasses and acquisition of the granule bulk surface images at varying bulk levels during granulation.
The vision system comprises of a high density illumination unit (a), a variable focus liquid lens optical system (b), a high-speed camera (c), and a liquid lens and an illumination unit controller (d, e).
In a laboratory setup, granule samples were subjected to rigorous mixing while being imaged at various distances, simulating the conditions in granulators
800 images of size 2048 px x 2048 px were acquired at a frame rate of 80 fps at each distance and for each granule size class.

Image analysis

Images of one size class (655.0 µm) were selected to represent optimally granulated particles. Optimal size class was determined based on the measured average particle size of the final product.
A convolutional neural network TriNet [2] was trained to recognize the visual appearance of the optimally granulated size class. The reference value for training was set to 1 for the optimally granulated size class and 0 for all the other size classes (binary classification).
On each test image, the optimally granulated regions were recognized: each output pixel of the prediction image is proportional to the probability of it representing an optimally granulated region.
The final prediction value is calculated as the mean of all prediction image values.

Evaluation

For each size class, the mean and the standard deviation of prediction values.
MAE from the reference values (the reference value is 1 for the optimally granulated size class and 0 for all the other classes).

Results and discussion

The results show that the trained neural network can recognize the specified optimally granulated size class, i.e., 655 µm, with high confidence. With a mean prediction value of the positive size class 0.99 (MAE 0.013) and a maximum mean prediction value 0.02 (MAE 0.020) within all negative size classes, the model achieved 100 % classification accuracy at threshold 0.5.

It is worth noting that the sole neighboring size classes, i.e., 550 µm and 780 µm, have a non-zero mean prediction value. Some regions in these images are detected as optimally granulated, which is expected.

In addition to providing a prediction value, the output of the neural net has the added benefit of a visual feedback displaying specific regions on the acquired images, where prediction values are high.

Figure 3. The major stages of the proposed approach

Conclusion

The results confirmed the feasibility of the deep learning-based approach for determination of the state of granulation within a controlled laboratory setup.

Our final goal is the application of the presented approach for the in-line and real-time determination of the optimal granulation endpoint, facilitating process development, scale-up and transfer, and increasing the product consistency during production in pharma.

See the full poster on “Deep Learning-Based Machine Vision System for Use in High-Shear Granulation” here

(click the picture to download the poster)

Source: Sensum poster “Deep Learning-Based Machine Vision System for Use in High-Shear Granulation”, Andraž Mehle, Domen Kitak, Domen Rački, Dejan Tomaževič, Sensum, Computer Vision Systems, Ljubljana, Slovenia; Laboratory of Imaging Technologies, Faculty of Electrical Engineering, University of Ljubljana, Slovenia

excipients formulation