Page 322 - IJB-8-4

Deep learning for EBB control
A B

Figure 3. (A) The flow diagram of the automatic parameter optimization algorithm. (B) An example of a 4-step optimization procedure. The
two steps, namely, the “init calibration” and the “small perturbations” steps, are highlighted, alongside example printability windows used
to calibrate the EM parameter. The dots in the figure correspond to the EM used for each step at the corresponding LH . Green dots represent
i
a print that was predicted as “ok” by the DL model, while red dots one that was predicted with an error (“under_e” or “over_e” classes).

viscosity, and elastic modulus) which are unknown in our model may show a high accuracy in testing, it may fail
case. to detect the correct quality of the print due to prediction
The optimization system starts with the EM uncertainty. To balance this problem, we introduce a
suggested by the initial printability window at the given second step called the “small perturbations” step in the
LH (Figure 3B, red dot in the first graph of the “init figure. After detecting a good quality print, we repeat the
i
calibration” step). Furthermore, at this stage, it also process by updating the EM by gradually smaller values
computes the difference between the two values for EM max (e.g., Δ/4 and Δ/8) until the DL model predicts “ok”
and EM evaluated at the specified LH Equation IV: again. The final optimized EM will be given by the mean
min value between the optimized EM of the “init calibration”

∆ = EM max (LH i ) − EM min (LH i ) (IV) procedure and the one from the “small perturbations” step.
To avoid wasting too much material and time, we
The Δ will be used as an update strength for the limit the overall number of updates (sum of the “init
EM during all steps in the optimization procedure. After calibration” and “small perturbations” steps) using a
printing, the DL model automatically evaluates the quality maximum iteration parameter (“max iters” in Figure 3A),
using the methods described in the previous sections. If the which can be decided by the user (default value equal
DL model prediction is an error, we update the printability to 5). To test the overall procedure, we printed a simple
window by making the first point the border point of one scaffold shape of 10 mm sides and 5 mm height using an
of the two corresponding cases (Figure 3B, green dot in infill density of 50% and a LH of 0.7. For this final test, we
i
the second graph of the “init calibration”). This is equal chose to use the transparent Pluronic solution, which was
to increasing (if the predicted class is “under_e”) or printed with the pneumatic-based extrusion bioprinter.
decreasing (if the predicted class is “over_e”) the EM of
the previous step by a factor of Δ/2. The process is then 4. Results and discussion
repeated until a good print is detected by the DL model. 4.1. Model optimization procedure
At the end of the “init calibration” step, a new
printability window has been defined and will not be Figure 4 reports the main results related to the model
updated in the following steps. However, even if the DL selection experiments. As shown in Figure 4A, all tested

314 International Journal of Bioprinting (2022)–Volume 8, Issue 4