Page 350 - IJB-9-4

International Journal of Bioprinting Bioprinting with machine learning

error, and the learning method is designed. The learning integrated learning algorithm itself is not a single machine
method is an easy logical processing of a layer of learners, learning algorithm, but through the construction and
instead of the results of weak learners. The results of combination of multiple machine learning machines to
training weak learners are input, and a learner is retrained complete the learning task, it has a high accuracy in the
to obtain the last output. In that case, we call the weak machine learning algorithm, with the shortcomings being
learner the primary learner, and the learner used for the complicated training process of the scheme and the low
combination is called the secondary learner. For the test efficiency .
[67]
sample, the primary learner is first predicted to gain the
input data of the secondary learner. Then, the secondary 4. Looking for more suitable bioinks
learner is predicted again to obtain the final prediction Bioprinted products are composed of bioinks that are
result . deposited layer by layer, and therefore, the performance
[70]
of bioinks has a significant impact on the final bioprinted
3.6. Comparison of the machine learning methods products. Due to the variety of bioprinting technologies,
The traditional machine learning algorithm represented it is difficult to find a unified standard to prepare bioink.
by K-nearest neighbor is generally considered shallow Advanced technology and high level of operators’
learning, and the features of the training data must be experience are often required to manually prepare bioink,
specified manually in the training process. Its feature and the quality of final product is not guaranteed. The
extraction ability of sample data and model generalization machine learning algorithm, combined with its powerful
is relatively weak, and it is usually applied in classification feature analysis ability, gives a perfect solution for the
and linear regression problems. Machine learning preparation of bioink through the configuration of bioink
algorithms represented by neural networks are considered in the existing cases and addressing the inconsistent quality
deep learning, which can automatically extract features of the final printed product.
from training data without human intervention. Therefore,
deep learning has better feature extraction ability and model The cell activity in bioink has a very important influence
generalization ability as well as good analysis effect on high- on the final printing effect. However, the low activity of cells
dimensional data, and can deal with nonlinear problems. in extrusion-based bioprinting limits its application. Reina-
[73]
However, the deep learning model has poor explanatory Romo et al. attempted to quantitatively analyze the effects
ability, involves many parameters, and consumes a lot of of cells in bioink based on machine learning algorithms.
resources in the training process. However, the traditional They studied the force of three hydrogels when passing
machine learning algorithm has a complete mathematical through two nozzles of different shapes. The collected
theory and strong interpretability. Moreover, the training data were analyzed by a machine learning algorithm called
of traditional machine learning algorithm models require Gaussian process. The algorithm was not only suitable
fewer resources . for estimating the importance of each parameter, but
[71]
Multi-layer perceptron, convolutional neural networks, also could calculate the influence of changing parameters
and long short-term memory neural networks all belong on the final printing results. The dataset was derived
to deep learning. When there are too many layers in the from the existing accurate data, and its adaptability was
multi-layer perceptron model, there will be too many verified by random partition. The effect of the framework
model parameters, resulting in high training costs and was verified by comparing the calculated results with
overfitting problems. Convolutional neural network the current experimental data. In order to maintain cell
solves these problems by using local connection and activity and withstand the pressure in the printing process,
[74]
parameter sharing to some extent. However, the data in the Allencherry et al. studied the configure method of
convolutional neural network can only travel forward and hydrogel and gelatin bath in extrusion-based bioprinting.
have no memory of the data that have been processed. Long A convolutional neural network was employed to identify
short-term memory neural network not only has memory the quality of the final prints. The objects processed by the
function, but also can deal with long-term dependence neural network and the data in the dataset were all images
problem. But the structure of long short-term memory obtained by optical microscope scanning, with a total of
neural network is too complex . 108 images in the dataset. A statistical analysis reported
[72]
that the accuracy of the model could reach 93.51%. Xu
Regardless of how a single machine learning model is et al. introduced a framework containing four machine
[75]
trained and optimized, it always has some shortcomings learning algorithms to determine viability of cells in
that are difficult to solve. Ensemble learning can combine bioink in stereolithography-based bioprinting. The four
a variety of machine learning algorithms to get a better algorithms are random forest, k-nearest neighbor, ridge
model to obtain better predictive performance. The regression, and neural networks. They first continuously

Volume 9 Issue 4 (2023) 342 https://doi.org/10.18063/ijb.739