2.1 Fabrication of the microfluidic device
Figure 1 shows the assembly of the microfluidic device. It has a poly(dimethylsiloxane) (PDMS)-made microchannel covered with a glass substrate and poly(methylmethacrylate) (PMMA) reservoir. A PDMS reservoir bonded to the PMMA reservoir separates the inside into two regions: cell culture media and a bicarbonate buffer for on-chip CO2 incubation, as shown in Fig. 1b.
A detailed description of device fabrication processes was described elsewhere[22]. Briefly, a master mold for a microfluidic channel was fabricated using backside photolithography[37]. Positive replicas with channel features were fabricated by molding PDMS (KE-106, Shin-Etsu, Tokyo, Japan) against the master mold. A 300 µm-thick PDMS membrane and the channel feature layers were treated by vacuum air plasma and assembled by direct bonding. The gel inlet and outlet were punched using a biopsy punch. A PDMS microchannel, a PDMS fitting fabricated from a 3D printed master mold, and a 26×76×0.5 mm glass slide were then plasma bonded. A channel-glass assembly, a PMMA reservoir (Proto Labs, MN, USA), a PDMS reservoir cast from a 3D-printed plastic mold, and a 3D-printed braille actuator fixture were assembled using adhesive.
2.2 Cell Culture
Human umbilical vein endothelial cells (HUVECs, C2517A, Lonza, Basel, Switzerland) were maintained in EGM-2 (Lonza), and cells at passages 5 to 6 were used for the experiments. Human lung fibroblasts (hLFs) (CC-2512, Lonza) were maintained in FGM-2 (Lonza), and cells at passages 5 to 7 were used. Both cells were maintained at 37°C and 5% CO2.
When the hLFs had grown and reached subconfluency in a T-25 flask, the medium was replaced with EGM-2, and the cells were incubated with CO2 overnight. The medium was then collected, and 0.2-µm-thick layers of cellular debris were removed. A 1:40 mixture of hLF-conditioned and fresh EGM-2 was used for microvascular culture.
2.3 Construction of the microvascular network
A matrix solution was prepared by dissolving 10 mg/ml (final) fibrinogen (F8630, Sigma‒Aldrich) and 0.2 mg/ml type I collagen (354236, Corning) in HEPES-buffered saline solution (HBSS, C-40010, PromoCell). HUVECs were detached and resuspended in the matrix solution to a final cell density of 8.0×106 cells/ml. The suspension was mixed with 0.2 U/ml thrombin (T4648, Sigma‒Aldrich) and loaded into the microchamber (See Fig. 1d, e). After allowing gel polymerization in a CO2 incubator for 5 min, the entire microchannel was primed with hLF-conditioned EGM-2. The PMMA reservoir was filled with 4 ml of the medium, and the PDMS reservoir with 1.5 ml of 0.8 M NaHCO3 with 65 mM Na2CO3.
The device was placed on a 37°C hotplate, and a braille cell (SC9, KGS, Saitama, Japan) was attached to the microfluidic device, and the cell culture medium circulated in the microchannel under the peristaltic actions of braille pins following a constant-flow rate waveform[23] of 0.25 Hz. To reduce pulsation caused by blank periods in peristaltic actions, we used dual 3-strand pumps (see Fig. 1d) driven half a cycle apart. The medium was circulated initially using the interstitial flow mode, as shown in Fig. 1e. The media and bicarbonate buffers were replaced every three days. After the route of the lumens connecting the two side channels was confirmed, the flow mode was switched to feedthrough flow mode by closing a valve using a braille pin. The direction of the feedthrough flow was changed every 24 hours.
The HUVECs in the microfluidic chamber were imaged using an inverted microscope (DMi8, Leica, Wetzlar, Germany) with a CMOS camera (DMK33UX174, The Imaging Source, Bremen, Germany). The images were 1920×1200-pixel grayscale and taken every 1 h or 1 d.
2.4 Image acquisition and processing
To prepare initial ground truth data, manual semantic segmentation of was performed on the 10 selected brightfield images obtained in the experiments. The authors empirically filled the 10 chosen images into RGB images in which three color channels correspond to the three classes (lumen, matrix, and lumen wall) on CLIP STUDIO Pro V2.0 (CELSYS, Tokyo, Japan). These ground truth image data were used to train a DeepLab3 + based on MobileNet-V2 in MATLAB R2022b Deep Learning Toolkit (MathWorks, Natick, MA, USA).
Using the above-trained network, semantic segmentation of the time-lapse images of the lumen was performed in MATLAB to obtain binary images representing the area occupied by the lumens. The change in lumen area was calculated from the sum of the ‘true’ pixels of these binary images. The images were then 2D meshed by im2mesh (https://www.mathworks.com/matlabcentral/fileexchange/71772). Then, FEATool V1.16 (Precise Simulation, Hong Kong) was used to calculate the velocity field and WSS inside the lumens.
2.5 Network extraction and analysis
The skeletonized lumen images and distance maps were obtained using standard functions of the MATLAB Image Processing Toolbox. The skeletonized images were separated into the nodes and edges using the same toolbox’s standard morphological operations. Then, the lumen region corresponding to each edge was reconstructed by inverse distance transformation of the distance map values masked by the edge pixels. The lumen region corresponding to each edge was used to mask the CFD result data value matrices and to extract the WSS and flow rate corresponding to the edge. The adjacency matrix was weighted with the WSS ratio \({R}_{\tau }\) or flow rate ratio \({R}_{Q}\):
\({R}_{\tau }={\sum }_{i\in \varvec{W}}{\tau }_{i}÷\frac{8\mu {U}_{\text{i}\text{n}\text{l}\text{e}\text{t}}}{\frac{{A}_{\text{l}\text{u}\text{m}\text{e}\text{n}}}{{d}_{\text{c}\text{h}}}} (\text{E}\text{q}.1)\) \({R}_{Q}=\frac{Q}{{U}_{\text{i}\text{n}\text{l}\text{e}\text{t}}{w}_{\text{i}\text{n}\text{l}\text{e}\text{t}}} (\text{E}\text{q}.2)\)
where \(\varvec{W}\) is the index set of pixels of the lumen walls corresponding to the edge, \({\tau }_{i}\) is the WSS at the position of the \(i\)-th pixel, \(\mu\) is the viscosity, \({U}_{\text{i}\text{n}\text{l}\text{e}\text{t}}, {w}_{\text{i}\text{n}\text{l}\text{e}\text{t}}\) is the velocity and width of the inlet of the culture chamber, \({A}_{\text{l}\text{u}\text{m}\text{e}\text{n}}\) is the area of the whole lumen, \({d}_{\text{c}\text{h}}\) is the distance between the pillars defining two perfusion channels, and \(Q\) is the flow rate at the lumen corresponding to the edge. Finally, the R 4.3.3 software environment and the igraph package (https://CRAN.R-project.org/package=igraph) were used to calculate the strengths, and betweenness.