Page 353 - IJB-9-4

International Journal of Bioprinting Bioprinting with machine learning

Figure 7. (A) Feature weight in random forest regression and (B) feature weight in random forest classification. Reprinted from ref. under the terms of
[87]
the Creative Commons CC-BY license.

The dataset was composed of 99 simulated data obtained the model could also optimize the printing parameters
by the simulation program. The results indicated that the while significantly reducing the number of experiments.
neural network with four hidden nodes in a hidden layer
had the best prediction accuracy, which could reach 90%. 6. Defect detection during 3D bioprinting
Fu et al. studied the optimization method of
[86]
parameters in extrusion-based bioprinting based on It is inevitable to have various problems in the 3D
a support vector machine algorithm. The parameters bioprinting process, which may lead to the final product
involved path height, nozzle temperature, nozzle size, and not meeting the relevant requirements. If the manual
composition. In addition to the optimization method for method is used to monitor errors in the printing procedure,
determining the printing parameters, the researchers also the method would be time-consuming and labor-
investigated the training effect of small-scale data sets on intensive, and require high level of skill and experience
machine learning algorithms. There were only 12 pieces of of the operator. The quality of 3D-bioprinted products is
data in the dataset they used, which were obtained in actual difficult to guarantee. Machine learning algorithms could
experiments. Using the support vector machine model automatically monitor the printing process and detect
created by them to evaluate the parameters in advance defects in time by automatically learning the characteristics
before extrusion-based bioprinting could ensure that the of existing defective products. Therefore, machine learning
probability of high-quality printing was higher than 75%. algorithms are also widely applied in the defect detection
In addition, they also studied the universality of the model. of 3D bioprinting.
Tian et al. collected data from published literature Jin et al. designed four models based on machine
[87]
[89]
and trained the machine learning regression model to learning algorithms to detect anomalies on each layer
find better parameters in extrusion-based bioprinting during 3D bioprinting. Support vector machine and
(Figure 7). The bioink contained alginate and gelatin, deep neural network algorithms are employed in the four
and the studied parameters included cell viability and models. A dataset including 240 images was constructed
extrusion pressure. Support vector regression and random by the team to detect defects in the transparent bioprinting
forest regression were utilized in the regression model, process and optimize the parameters of the printing
and the dataset contained 956 data. The calculation results process. Experiment results revealed that the model with a
verified that the model could not only predict the printing conventional neural network worked best. They expected
effect in the original literature after a lot of training, but that introducing transfer learning into the model would
also further analyze the details. The research results once help reduce the burden of building a dataset and improve its
again confirmed the effectiveness of machine learning in versatility. In order to find out the possible quality problems
the design of bioprinting experiments. Ruberu et al. in the extrusion-based bioprinting process in time, Bonatti
[88]
investigated the printability of extrusion-based bioprinting et al. added a long short-term memory neural network
[90]
based on the Bayesian optimization algorithm in a search to the bioprinting control process (Figure 8). A camera was
for better printing parameters. Unlike other machine placed in front of the printer to film the printing process
learning algorithms, the Bayesian optimization algorithm and extract the dataset for training the neural network.
did not need a lot of data to train the model. Its advantage The team constantly modified the printing parameters to
was that the construction of the architecture was completed obtain videos of the bioprinting in different states, which
with only a few experimental data. The experimental ultimately improved the comprehensiveness of the dataset.
results showed that compared with the previous methods, Experimental results showed that the method could

Volume 9 Issue 4 (2023) 345 https://doi.org/10.18063/ijb.739