Page 254 - v11i4

International Journal of Bioprinting Deep learning-based 3D digital model of fetal heart

the segmented area that contains the entire structure of prints with smooth surface finishes, which are essential for
the heart, further generating the position mask. The mask accurately reproducing the intricate anatomical details of
is a matrix whose shape and position directly correspond the fetal heart, such as chambers and vessels.
to those of the image. To ensure that all subsequent A key characteristic of the Stratasys J750 system is
operations are within the scope of the mask, all elements that fundamental print parameters, unlike those in many
of the mask within the segmented area were labeled 1,
while other elements were labeled 0. The restricted area FDM or SLA systems, are largely controlled internally and
was obtained by multiplying the mask with the original are not user-adjustable. Instead, print characteristics are
image; the grayscale distribution of the original image and primarily determined by the selection and combination
the restricted area was calculated and plotted as curves. of photopolymer materials. For this study, a blend of
Sonographers could select an appropriate threshold to flexible, rubber-like material and rigid resin was chosen.
segment the restricted area by relying on the curves. By This material combination resulted in models with a Shore
repeating the operation for each image decomposed from D hardness of approximately 60, providing a balance of
the echocardiographic volume data, the segmentation of structural stability and pliability. The models were printed
scan data was obtained. Finally, the smoothing algorithm at a high-quality setting, achieving a layer resolution of
and seed fill algorithm were employed during the post- 200 µm. These properties enhance the models’
processing of segmentation. 21 utility for clinical visualization, education, and
diagnostic communication.
The fetal heart is a continuous solid body. Therefore, if
the interval between adjacent layers of the image scanned 2.7. Evaluation
by ultrasonic instrument (d) is small, it can be assumed To evaluate the performance of the segmentation task,
that the images and their segmentations are similar the labels from the test set were used as the standard to
between layers, or at least not significantly different. An calculate the positive prediction rate, recalling rate, mean
end-diastolic volume data consists of coronal, axial, and intersection over union (mIOU), and Dice similarity
sagittal raw sectioned images. Axial raw sectioned images coefficient (DSC). According to the standard label (sl) from
were generated at 0.2 cm intervals and were sent to FRT experienced sonographers and segmentation result (sr),
for segmentation. Surface reconstruction of the entire the pixel can be divided into four categories: true positive
structures of the fetal heart was simultaneously achieved (TP; sl: 1, sr: 1), false positive (FP; sl: 0, sr: 1), true negative
by stacking the outlines of the segmentation results layer (TN; sl: 0, sr: 0), and false negative (FN; sl: 1, sl: 0). For
by layer. (x, y, z) is the coordinate of a point on the model, the segmentation result of a single ultrasonic image, the
and model (x, y, z) indicates whether the point is part of the values of TP, FP, TN, and FN are the total number of pixels
model; 1 denotes “yes” and 0 denotes “no.” The concept of of the corresponding category, respectively. Therefore,
similarity, dst (x, y) indicates whether the point is part of the the positive prediction rate (P) and recalling rate (R) are
i
model on the 2D segmentation result of layer i from FRT. defined as:
Therefore, the mathematical representation of the model is:
TP
model (x, y, z) = dst i (x, y) (I) P (II)
TP FP

where, in normal circumstances, z = i.
Removal of free structure and surface smoothing R TP (III)

was performed in Mimics software during model post- TP FN
processing to obtain a smoother and more intuitive
digital model. Thereafter, the 3D model was printed According to the prediction rate and recalling rate, the
with an SLA 3D printer (J750, Stratasys, USA) using mIOU and DSC can be defined as:
photosensitive material.
TP
2.6. 3D printing process mIOU (IV)
The 3D digital fetal heart models were fabricated into TP FN FP
physical models using a Stratasys J750 PolyJet 3D printer
(Stratasys, USA). This multi-material inkjet technology TP
was selected for its capability to achieve high-resolution DSC (V)
TP FN TP FP

Volume 11 Issue 4 (2025) 246 doi: 10.36922/IJB025200192