Page 184 - IJB-8-1

Analyzing Cell-Scaffold Interaction through Unsupervised 3D Nuclei Segmentation
A B C

D E F

Figure 3. (A) Schematic diagram of scaffold with fiber stacking structure. (B) scanning electron microscope images of overall fibrous
scaffold structure. (C) Cross-section of fiber stacking. (D-F) Fiber surface morphology of poly-E-caprolactone (PCL), PCL-10-D and PCL-
20-D scaffold, respectively. Figure 3(A)-(D) are original images, and Figure 3(E) and (F) are adapted from ref. licensed under Creative
[20]
Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0), https://creativecommons.org/licenses/by-
nc-nd/4.0.
These scaffolds are used to culture NIH-3T3 mouse developed AD-GAN model to output non-overlapping
embryonic fibroblast cells, A549 human non-small 3D nuclei. The probability of rotating directions of
cell lung cancer cells and HeLa cells. The nuclei and ellipsoids is assumed the same in all directions to mimic
membrane of the three cell lines are stained with Hoechst real nuclei, which encourages the AD-GAN model to
33342 (blue) and DiI (red) for CLSM imaging. The living- map Z-axis elongated nuclei caused by light diffraction
image of cell-seeded scaffolds after fluorescent staining and scattering in real CLSM images to the ellipsoids in
are taken by CLSM (LSM-880, ZEISS, Germany) with the synthetic masks. To achieve one-to-one mapping
an EC Plan-Neofluar 20X/0.5 air immersion objective. between real images and synthetic masks, the proposed
The CLSM images are compiled using Z-stack mode by AD-GAN training includes both same-domain translation
ZEN software, and reconstructed using Imaris software (image-to-image and mask-to-mask) and cross-domain
(Bitplane Inc). The collected CLSM images have been translation (image-to-mask and mask-to-image) as shown
resized to 512×512×64 voxels by spatial normalization to in (Figure 4B). A single GAN-based auto-encoder is
remove redundant information (data available at: https:// designed to build a bidirectional mapping between each
github.com/Kaiseem/Scaffold-A549). The processed real image and the corresponding content representation
images are split into 16 image patches with the size of in the shared content space.
128×128×64 voxels as the inputs of AD-GAN model. Our proposed AD-GAN model consists of a unified
conditional generator and a PatchGAN discriminator Ɗ,
2.2. AD-GAN method and training strategy both of which use 3D convolutional layers. As shown
Unsupervised nuclei segmentation can be considered an in (Figure 4A), the designed generator in each domain
image-to-image translation task as shown in Figure 4, contains two parts: an encoder (ꞔ ) which encodes the
enc
where the inputs are the grayscale CLSM images, and input volume to content representation, and a decoder
the output images are the segmentation results. Our AD- (ꞔ ) which reconstructs the content representation to
dec
GAN model is essentially designed to deal with two the output volume. In the same-domain translation, the
domains: domain A with image style including input generator is trained to extract useful information by auto-
image, reconstructed image and cyclic image, and domain encoding. In the cross-domain translation, the decoder ꞔ
dec
B with mask style including input mask, reconstructed is frozen and the encoder ꞔ is trained to generate fake
enc
mask and fake mask. The synthetic mask is generated images to fool the discriminator by aligning the content for
by non-overlapping ellipsoid structures with random each domain. The encoder contains two down-sampling
rotations and translations, which would stimulate the modules and four ResNet blocks, and the decoder has a

170 International Journal of Bioprinting (2022)–Volume 8, Issue 1