Page 59 - IJB-9-4

International Journal of Bioprinting Machine learning and 3D bioprinting

many features can be extracted based on the expert users’ flexibility in discovering hidden patterns and relationships in
prior knowledge and practical experience. For numerical complex images. For both methods, the parameters could be
data, statistical analysis, threshold methods, and frequency adjusted to improve the classification and regression models.
analysis are typically used for feature extraction. By contrast, Labeled/unlabeled datasets, dataset size, and task complexity
shape, edge, color, and texture detection are widely used are key factors in ML method selection.
for image feature extraction. To achieve better predictions, Researchers worldwide have explored ML applications in
the informative relevance of these features should be bioprinting from fundamental perspectives to performance
carefully chosen for each task . Traditional ML methods, modeling. The potential adoptions of DL in design and
[20]
such as Support Vector Machines (SVM) and K-Nearest fabrication of patient-specific 3D tissue-engineered
Neighbors, are typically used to build classification or constructs were reviewed, such as image-processing and
regression models for such prediction tasks. For example, segmentation, optimization and in situ correction of printing
classification models can classify Taylor cone shapes in parameters and refinement of the tissue maturation process .
[5]
EHD bioprinting, identify the reliability of printability, The authors also reported some relevant practical applications
and judge the shape fidelity. Regression models can be and summarized that the availability of huge training datasets
used to optimize the printing parameters and material and well-defined evaluation metrics are the key factors to
properties for the desired fiber diameter. One of the most accelerate the corresponding research areas.
commonly used ML methods is SVM, and a few powerful
[23]
SVM toolboxes have been launched for prediction models Shin et al. discussed the supervised ML, unsupervised
built in MATLAB or other platforms with user-friendly ML, semi-supervised ML and reinforcement ML methods,
interfaces and simple instructions . and their applications in preprinting, printing, and
[21]
postprinting. They concluded that ML can optimize printing
As illustrated in Figure 3b, DL methods can parameters and bioinks, save printing time, and detect
automatically discover underlying patterns and identify the the anomalies. The identified bottlenecks are the limited
most descriptive and salient features in image-recognition amount of data and the transferability of current models,
tasks. Feature extraction and selection steps are omitted in since these models heavily depend on mathematical features
this type of learning method. DL methods have achieved of the training data and may suffer from inconsistency when
significant success in many image application scenarios, dealing with data from other sources.
and convolutional neural networks (CNN) are the most
popular DL methods . Unlike traditional ML methods, The aim of this work is to collect and summarize the
[15]
DL methods can handle more complex tasks, such as publications and present a state-of-the-art review, and
classification, segmentation, and object detection, which highlight the emerging scientific potential of ML when
require relatively large-scale datasets. applied to bioprinting. First, ML applications in process
monitoring, printing parameters, and biomaterial/
Traditional ML and DL methods used in bioprinting bioink design are discussed. Second, ML applications in
applications can also be categorized into supervised and cell performance analysis are investigated. After that, a
unsupervised learning . Supervised ML methods can literature-based analysis of challenges and outlook is given.
[21]
establish mathematical models between inputs and outputs
using labeled datasets, such as multilayer perceptron 2. ML applications in process, parameter,
(MLP) , SVM [22,25,26] , CNN [15,22] , and backpropagation and material optimization
[24]
neural network (BPNN) . Because of the complicated
[27]
nature of images, some objects are difficult to label, such The published papers on ML applications in bioprinting
as nuclei, morphological phenotypes, and cell shapes are summarized in Table 1, along with the application
during proliferation and migration. Unsupervised ML areas, specific tasks, and proposed ML algorithms.
methods have been proposed to explore unlabeled objects Current applications are divided into four categories:
as patterns or clusters and to identify hidden patterns or image analysis-based in situ process monitoring, printing
similarities through self-taught rules [28,29] . It was believed parameter optimization, biomaterial/bioink optimization,
that ML applications in process optimization, dimensional and cell performance analysis. Various properties are
accuracy analysis, manufacturing defect detection, considered for parameter optimization such as fiber
and material property prediction may accelerate the resolution, integrity of the printed constructs, cell viability
perspectives of bioprinting development . after extrusion and the integration of them. Even though
[22]
diverse tasks are listed, the focus is on building process–
In short, traditional ML methods are fully transparent, material–performance models. It can be seen that both
and researchers can transfer their insights and knowledge to traditional ML and DL methods can be applied, and the
domain-specific tasks. DL methods are superior in terms of ML method selection often considers the task complexity

Volume 9 Issue 4 (2023) 51 https://doi.org/10.18063/ijb.717