Page 64 - IJB-9-4

International Journal of Bioprinting Machine learning and 3D bioprinting

the advantages of multiple ML methods, this learning tree regression, and k-nearest neighbor (KNN) regression,
algorithm used the weighted average of multiple ML models were developed and compared. The RFr model yielded
to improve its robustness. This model demonstrated a high the highest accuracy when predicting the proliferation
prediction accuracy under various printing conditions. rate of L929 fibroblast cells after a 7-day culture, and this
Although the necessity of this cell viability study is high, prediction model also coincided with the data collected
the available datasets for this research topic are limited . from in vivo studies. This demonstrates the potential
[46]
This is in part caused by the tedious and expensive cell of ML for discovering the impact of fiber diameter on
viability data collection process. angiogenesis. However, the generalization performance of
DL methods, such as CNN, can advance cell viability the developed model has not been verified for more cell
studies by segmenting cells directly in fluorescence types and scaffold structures.
images, which can rapidly screen the images with less Traditional ML algorithms, including KNN, logistic
manpower and effort. A dataset comprising images of regression, RFc, SVM, and ANN, have also been utilized
4974 single spheroids with corresponding labels was to evaluate delta permittivity in a scaffold-based cell
used to build the prediction model. The developed CNN culture environment. Four cell types were cultured on the
model successfully categorized cell viability into three PCL scaffolds, and the corresponding delta permittivity
classes with a balanced accuracy of 78.7%. CNN has was measured over time using dielectric impedance
also demonstrated good generalization performance spectroscopy to build the dataset. Input features were
when predicting the cell viability of bioprinted renal extracted from the relative permittivities at different
spheroids under varied inhibitory concentrations as well frequencies, time points, and cell types . Among the
[50]
as experimental settings . This may be a new pathway to developed ML models, KNN yielded the best accuracy in
[47]
prompt efficient cell viability studies by segmenting cell/ classifying cell types and culture durations.
nuclei in fluorescence images and then counting live/dead Biological images, such as confocal laser scanning
cells using the developed DL models. Nevertheless, cell microscopy (CLSM), are widely used to visualize cell
assays using multidimensional fluorescence images are behavior and reveal cell–microenvironment interactions .
[38]
required for more comprehensive and accurate analysis.
However, the cell shape and morphological features of these
images were difficult to extract using statistical methods.
3.2. Cell–microenvironment interaction ML algorithms have been proposed to quantitatively
Cell–microenvironment interactions are crucial for translate and deliver this information to researchers. SVM
immune response and tissue regeneration. It is well known can quantify the diverse cell morphologies of hBMSCs
that the cell response varies with both materials and their populations using CLSM images and associate them
forms. Even for the same material, the cell response may with specific microenvironments, such as PCL fibrous
vary significantly when interacting with nanoparticles, substrates and PCL spin-coated films . SVM can also
[51]
scaffolds, coatings, or films. Some cell types may detect the impact of bioprinted construct morphology
experience the benefits/risks associated with particular on cell-shape phenotypes . Using the extracted metrics
[35]
material compositions or forms. To investigate the needs from cellular and subcellular morphometry, the developed
and preferences of cell growth in bioprinted constructs, SVM model successfully classified the substrates into
their behavior should be digitalized for in situ analysis.
either woven PCL mesh with precision-stacked microscale
There has been a growing interest in applying ML fibers or nonwoven mesh with randomly oriented fibers.
to identify cell types, phenotypes, and shapes. ML has As pioneering research has linked cellular and subcellular
outperformed experts in segmenting and classifying morphometry with substrate topology, the quality of
cell/nuclei in biological images across various tasks [48,49] . the training dataset is critical for reliable identification.
Motivated by this progress, researchers have initiated However, the size and diversity of the datasets used in the
studies applying ML for cell–microenvironment aforementioned studies were not mentioned. This could
interaction analysis, that is, cell–scaffold interaction, cell– raise concerns regarding the robustness and convergence
cell interaction, and cell–material interaction [35,36] . of the proposed SVM models when dealing with images
Traditional ML methods can model the relationship collected from different substrate morphologies.
between cell proliferation and the physicochemical Owing to the complicated nature of biological images,
properties of electrospun scaffolds with regard to the fiber it is difficult to prepare datasets with labeled nuclei,
diameter, pore diameter, water contact angle, and Young’s morphological phenotypes, or labeled cell shapes for
modulus . Six regression algorithms, namely, linear proliferation and migration. Therefore, unsupervised
[50]
regression, SVM regression, RFr, lasso regression, decision methods have been applied to model the objects of interest.

Volume 9 Issue 4 (2023) 56 https://doi.org/10.18063/ijb.717