Open Access
Add to BibSonomy
Abstract
Two-photon polymerization (TPP) is an advanced 3D fabrication technique capable of creating features with submicron precision. A primary challenge in TPP lies in the facile and accurate characterization of fabrication quality, particularly for structures possessing complex internal features. In this study, we introduce an automated brightfield layerwise evaluation technique that enables a simple-to-implement approach for in situ monitoring and quality assessment of TPP-fabricated structures. Our approach relies on sequentially acquired brightfield images during the TPP writing process and using background subtraction and image processing to extract layered spatial features. We experimentally validate our method by printing a fibrous tissue scaffold and successfully achieve an overall system-adjusted fidelity of 87.5% in situ. Our method is readily adaptable in most TPP systems and can potentially facilitate high-quality TPP manufacturing of sophisticated microstructures.
© 2024 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
Direct laser writing (DLW) by two-photon polymerization (TPP) has emerged as a revolutionary additive manufacturing (AM) technique for micro- and nanofabrication [1]. It utilizes two-photon absorption (TPA) induced in the small focal region of a high-intensity, ultrafast laser to initiate bonding of the monomers in the photoresists. The unit volume of polymerization is below the optical diffraction limit, offering TPP high spatial resolution compared to other freeform 3D printing technologies. TPP has been used as a prototyping tool for metamaterials, microfluidic devices, miniaturized optics and tissue scaffolds [2–5], among other creative applications.
Similar to other AM processes, the experimental parameters for TPP fabrication are often optimized to ensure the initial design achieves its ideal surface shape. However, sophisticated microarchitectures such as biomimetic scaffolds have intricate embedded features [6,7], yielding the need for visualizing internal patterns of 3D printed parts. Characterization methods like X-ray microscopy [8] and confocal microscopy [9] have been employed, but they are limited by the time-consuming step of washing out unpolymerized photoresists. Defects can arise during the writing process due to several reasons, including the proximity effect and thermal damage [10,11]. In such circumstances, the fabricated structure is often not inspected until after it has been developed, usually in a cleanroom. Thus, this overall production cycle—design to fabrication to analysis to final product—is characterized by a slow and costly post-fabrication process, which hampers efficiency, optimization and innovation.
In situ inspection of TPP-fabricated structures offers the opportunity to assess fabrication quality and subsequently adjust the writing conditions in real time. To date, there have been a few approaches incorporated to address this challenge. For instance, researchers used optical coherence tomography to inspect post-fabricated structures [12]. Additionally, optical diffraction tomography [13] and broadband coherent anti-Stokes Raman scattering microscopy [14] are utilized in other cases. In all of these examples, separate light sources are required and the inspection system needs to be constructed alongside the fabrication platform, thereby increasing cost and complexity, particularly for the co-alignment of two separate instruments. A promising alternative utilizes multiphoton imaging, namely either third-harmonic generation (THG) microscopy [15] or two-photon excited photoluminescence [16] for in situ inspection. The appeal in these cases is that these imaging modalities share the same ultrafast excitation laser used for TPP. However, the tradeoff is that THG is only sensitive to interfaces between media, i.e., refractive-index contrast, which makes in situ inspection very design specific. Additionally, inspection by multiphoton imaging relies on endogenous fluorescence, the strength of which poses restrictions on the materials (photopolymers) used for fabrication. Furthermore, the speed of these techniques is generally limited by point-scanning and can only be incorporated after polymerization.
Here, we demonstrate an alternative approach to enable in situ monitoring during TPP fabrication that leverages the speed afforded by full-frame widefield detection and the simplicity of standard brightfield (BF) illumination that accompanies almost all TPP systems. This approach presents several advantages, including a high data-acquisition speed, real-time feedback during the writing process, and no constraints on the type of photoresist used. Until now the primary concern of using BF microscopy for inspection during TPP fabrication has been that it only provides a two-dimensional overview of the polymerized regions, lacking the necessary optical sectioning to adequately capture the 3D intricacy of the structure. Machine learning (ML) models trained by video data collected under BF have achieved automatic detection of TPP part quality [17], but it is computationally expensive and can only classify certain types of failures.
We introduce a simple and generic approach, termed automated brightfield layerwise evaluation (ABLE), which utilizes BF microscopy for in situ and real-time quality assessment during TPP fabrication, without the need for ML tools. During the fabrication process, layerwise images are captured and background subtraction is employed to highlight the spatial characteristics of each layer. We also employ an estimated impulse response of the ABLE-TPP system, along with the input design, to predict the fabrication outcome for precise quantitative analysis. We find that our approach provides a 3D visualization of the TPP fabrication process and enables automated assessment of defects and deformations without any additional hardware modifications. Furthermore, we demonstrate that our method achieves an average system-adjusted fabrication fidelity of 87.5% for printed fibrous tissue scaffolds derived from bovine tendon tissues, while adding only 34 ms to the data acquisition time per layer.
2. Methods
2.1 System workflow
We present a framework for process monitoring and quality assessment of TPP manufacturing based on brightfield imaging, system analysis and point-set registration. Taking advantage of the unique characteristics of layerwise AM processes [18], we demonstrate an automated program with defect and deformation detection capabilities, thereby permitting on-demand troubleshooting during the writing process.
Figure 1 describes the workflow diagram of the proposed ABLE-TPP system. The top row of the diagram illustrates the data flow throughout the process. First, the program sequentially reads a binary image stack Jn(x,y) as mask patterns that guide the fabrication of each layer (e.g. 3D models are sectioned into 2D images), where x and y represent horizontal and vertical spatial coordinates, respectively. In this context, the subscript n denotes the layer number, and each J(x,y) represents a binarized 2D matrix comprising the spatial information of the masks. Jn(x,y) are convolved with the system’s pre-measured impulse response H(x,y), to create reference templates Tn(x,y) for post-exposure spatial-feature analysis. Next, the TPP platform replicates the design using the same processing parameters as in H(x,y), while a BF imaging system concurrently monitors and records the process as an intensity map, denoted as In(x,y).
Fabrication progresses from the bottom to the top layer; this approach enables the acquisition of spatial intensity distributions of the preceding layers, which is then leveraged to extract single-layer information Fn(x,y). In the final stage, Fn(x,y) and Tn(x,y) undergo filtering and binarization, followed by a comparison to evaluate the fabrication quality. This comparison yields a system-adjusted fidelity parameter, denoted by q, which is calculated as the percent spatial overlap between the two sets of binary images. A user-defined threshold qthr determines whether to proceed or stop the fabrication. A 3D fidelity map is generated at the end of fabrication, and a fidelity matrix is also available for running multiple designs with different processing parameters. The detailed steps are explained in the following sections. Note that for brevity we henceforth drop the explicit dependence on the spatial coordinates in our notations.
2.2 TPP fabrication
2.2.1 Materials
The photoresist is composed of 10% w/v methacrylated gelatin (GelMA, Advanced BioMatrix) in phosphate buffered saline (PBS) as the prepolymer and 4.1 mM Rose Bengal (Sigma Aldrich) as the photoinitiator. We dropcast 100 µL of photoresist on a coverslip previously cleaned with ethanol and dried. A second coverslip is placed atop the resin to prevent dehydration, and is supported by a 1.19-mm thick spacer mounted between the two coverslips for consistent separation.
2.2.2 Photochemistry
TPP is a direct consequence of TPA, a third-order nonlinear optical process where a ground-state electron simultaneously absorbs two photons to transcend the energy gap to reach a first-excited electronic state. TPA is typically adapted under the so-called degenerate condition using two photons of identical frequencies, but it could also be achieved with photons of different frequencies for the non-degenerate case. The probability of the photoinitiator to undergo TPA depends quadratically on the laser intensity, offering TPP a much higher axial resolution than single-photon polymerization (see Appendix). Upon light excitation, the photoinitiator generates singlet oxygen that initiates a chain-growth radical polymerization of the methacrylamide and methacrylate side groups on the GelMA chains, eventually forming a crosslinked network of gelatin polymers.
2.2.3 Optical setup
The schematic setup for our TPP platform is shown in Fig. 2. An ultrafast laser source (InSight X3, Spectra Physics) generates a beam at 800-nm wavelength, 120-fs pulse width and 80-MHz repetition rate, followed by a half-wave plate and a linear polarizer to adjust the output laser power. The beam is coupled to an inverted microscope (IX83, Olympus) and tightly focused by an oil-immersion 40X/1.3 NA objective (Olympus) into the photoresist. The volume of the photoresist that is solidified in a single exposure is referred to as a voxel. The fabrication occurs in a layer-by-layer manner, with a dual-axis scanning galvo mirror system (GVS002, Thorlabs) directing the transverse (x and y) movement of the laser beam and a motorized stage to move the objective in the axial (z) direction, where the mechanical resolution is 15 µrad for the galvo and 100 nm for the z stage. The distance between two voxels in the x-y plane is defined as the hatching distance Δx, and the distance between two voxels in z direction is defined as the slicing distance Δz. The average laser power used in this work is 36.7 mW at the focus, and the writing speed is 0.2 mm/s. The microscope is equipped with a broadband light source for BF illumination. The objective is used for both focusing the laser for printing and collecting the BF images. A tube lens (f = 180 mm) relays the focal plane to a scientific complementary metal-oxide-semiconductor (sCMOS) camera (C15440-20UP, Hamamatsu). To improve the contrast of the images, we utilize a 680-nm shortpass filter to block the direct laser light.
Fig. 2. Schematic setup of the TPP system (L1, L2: lens; PH: pinhole; LP: linear polarizer; HWP: half-wave plate; SL: scanning lens; TL: tube lens; DM: dichroic mirror; Obj: objective; CL: condenser lens; M: mirror; SPF: shortpass filter). Close-up view of voxel-by-voxel printing in TPP (Δx: hatching distance; Δz: slicing distance).
2.3 Process monitoring and data acquisition
The BF imaging system facilitates real-time monitoring and data collection of the fabrication process. However, conventional BF microscopy has 3D ambiguity due to the blurry appearance of out-of-focus features. To increase the optical sectioning ability, we introduce adaptive background subtraction. We refer to the layered information currently under fabrication as ‘foreground’ and the optical contributions from other layers as ‘background’. When the objective stage moves to a new axial location, the camera is triggered to take a snapshot and a background image Bn is obtained. After the fabrication is completed in this layer, another image In is captured before the stage moves to the next location. By subtracting In with Bn, the foreground image Fn can be determined. This operation is executed continuously during the fabrication process, with each layered information registered with its axial location. The camera parameters can be adjusted in our customized program, with the exposure time set to be 17 ms and the window size equal to 541 × 541 pixels in our case.
The contrast of Fn stems from the local refractive index change induced by polymerization, i.e., dark features (polymerized) on a bright background (unpolymerized). To always see a sharp and focused image when the objective moves in z, the camera position is adjusted so that the conjugate plane of the sensor matches with the laser focus. The process of aligning the camera is depicted in Fig. 3. First, we select a central axial plane z0 and capture its corresponding B0. Next, we move the stage 5 layers both below and above z0 at an axial distance of 100 nm. During this process, straight lines created from a single scan of 70.3 µm are printed on each plane at varying y coordinates, spaced 7.0 µm apart. Once the printing is complete, the stage is repositioned back to z0 to capture I0. We then evaluate the line intensity profiles in F0 and fine-tune the camera’s position to ensure that the central line at z0 exhibits the highest contrast. Following calibration, the contrast of the central line increased from 0.46, as shown in Fig. 3(a), to 1.94 in Fig. 3(b).
Fig. 3. F0 and the intensity profiles along the red dashed lines (a) pre-calibration and (b) post-calibration.
2.4 Quantitative analysis
2.4.1 Generating reference templates
Reference templates, denoted as Tn, are generated for direct comparison with the outcomes of the ABLE-TPP system. This comparison is facilitated through the convolution of the mask pattern Jn with the system’s impulse response H, effectively simulating the impact of the entire printing-to-imaging workflow on the input design.
How do we derive H? The ABLE-TPP system operates as an integrated platform that combines 3D microfabrication capabilities with subsequent imaging of the fabricated structures. From a system and signal perspective, this process involves transforming a digital signal (design file) into a physical output (3D printed object) and then analyzing this output through an imaging process to generate another form of digital signal (image data). Similarly, H is twofold: firstly, it represents the physical manifestation of the impulse signal as a printed voxel; secondly, it embodies the imaging system's point spread function when capturing an image of the voxel.
Under the same printing parameters and illumination conditions, we assume the system to be linear and shift-invariant (LSI), meaning that the image of voxels will have the same appearance at any spatial coordinate, and that the overlapping of voxels results in a superposition of their pixel intensities. In reality, the ABLE-TPP system does not strictly adhere to the principles of an LSI system due to several factors. For example, there is a diminution in laser power towards the periphery of the printing field-of-view (FOV). Additionally, the system is susceptible to error from the proximity effect in instances of high pattern density. However, for the purposes of simulation, ABLE-TPP is considered to operate within a theoretical framework where such non-ideal effects are absent. This assumption establishes a benchmark for expected printing quality, representing an idealized outcome against which actual products can be compared. It also serves to provide a practical scope for the experimental parameters.
By printing and imaging a dot grid as a two-dimensional impulse train, H is retrieved using
Thus, Tn is acquired by
2.4.2 Data preparation
The recorded foreground image data Fn is further processed in MATLAB to optimally extract the key features. A bandpass filter is applied in Fourier space to filter out large structures down to 40 pixels and small structures up to 3 pixels. The specified size is chosen to effectively reduce noise and smooth the shading in the images. IsoData thresholding [19] is then applied to identify significant differences that likely represent foreground objects, resulting in binary image data set Fn*. The corresponding templates are also filtered and binarized using the same approach, resulting in Tn*.
Before comparing Fn* and Tn*, we match the coordinate systems of the two image stacks by using a random sample consensus (RANSAC) [20] algorithm iteratively to minimize the outliers in the matching SIFT (scale invariant feature transform) features [21]. Here we use the sum of squared differences between matches to calculate the similarity scores. Once a satisfactory plane fit is identified for the first layer (i.e., when the similarity score is the smallest), a transform matrix is calculated and applied to align the subsequent layers.
2.4.3 Fidelity calculation and error detection
The fabrication quality is assessed using system-adjusted fidelity q, which measures the part’s geometric conformity with an ‘adjusted’ design that factors in the system characteristics. q is determined through a pixel-by-pixel comparison to calculate the spatial overlap between Fn* and Tn*, with normalization based on the total pixel count in the printed area. A high q value suggests a superior fabrication outcome, indicating that key features are accurately reproduced. Conversely, a low q value signifies a less precise replication of the desired features. In addition, a user-defined threshold qthr is implemented to stop the writing process if the fidelity falls below an acceptable level.
A grid-based analysis is used to study the spatial distributions of q. Fn* and Tn* are merged as different color composites and segmented into uniform square grids. Within these grids, local q values are calculated. Grids lacking expected features are categorized as dark regions and excluded from analysis. For grids that have q values below threshold, they are scrutinized for potential defects, marking area for further analysis and process refinement. Upon completion of the entire structure, we calculate the probability of errors across all layers for each grid. This probability is normalized against the instances where the grid does not constitute a dark region. Finally, the program outputs q values for the non-dark regions in each layer, displaying these values along the z-axis.
3. Results and discussion
Figure 4 shows the process of characterizing the system’s impulse response H. With the selected writing speed and laser power, a dot grid of 7 by 7 with horizontal and vertical spacings of 5.0 µm was printed. The spacings are much larger compared to the lateral dimension of a single voxel, allowing the impulse signals to be considered as independent. The foreground image F of the dot grid is depicted in pseudo-color and presented in Fig. 4(b). The top right inset displays H derived based on regression analysis over all the dots in the image. We observe that the shape of H has an ellipticity of 0.58, which is likely a result of the astigmatism in the optical setup.
Fig. 4. (a) J of a 2D impulse train. Pixels valued at 1 indicate the regions designated for printing. (b) F of the corresponding fabrication result with the inset shows the best estimation of H from regression analysis. The color bar represents the pixel intensity in an 8-bit image.
We next present the inspection of a fibrous tissue scaffold fabricated by our ABLE-TPP system. The design is derived from second-harmonic generation (SHG) image data of bovine tendon tissue with 60 optical sections taken at 1 µm apart. Figure 5(a) shows the SHG image at layer 11 counting up from the bottom. Ridge detection was applied to generate mask patterns Jn with major fiber networks in order to discern individual fibers in the binarized SHG images [22], as shown in Fig. 5(b). Next, the reference templates Tn are generated by convolving H with Jn in which pixels are spaced at a predefined Δx. When Δx = 279 nm, the generated template is shown in Fig. 5(c). It is considered a successful outcome given voxels overlap properly, resulting in continuous fibers that remain distinguishable from one another. In other cases, a tighter Δx of 0.216 µm produced fibers with a thicker diameter and the formation of undesirable joint regions, as can be seen in Fig. 5(d). Similarly, a larger Δx can give rise to disconnected fibers (not shown here). These results show that our method provides an easy solution to find an optimal processing parameter space prior to fabrication.
Fig. 5. (a) The original SHG image of the collagen from a bovine tendon tissue at layer 11 (scale bar: 70 µm). (b) J11 generated from ridge detection. Simulation results when Δx is (c) 0.279 µm and (d) 0.216 µm.
The 3D rendering of the fibrous scaffold is displayed in Fig. 6(a). The scaffold possesses intricate internal features, which exhibit overlapping in the z direction. We first examine the effectiveness of adaptive background subtraction in isolating features of a single layer. Figure 6(b) and Fig. 6(c) are the brightfield images I directly captured by the camera on layers 5 and 25, respectively. Features from underlying layers are evident in both images, which obscure the visibility of the layers of interest. Conversely, the corresponding images with subtracted backgrounds F, as illustrated in Fig. 6(d) and Fig. 6(e), effectively retain in-layer details while eliminating out-of-layer contributions, unveiling single-layer information for both observation and analysis. These results confirm that adaptive background subtraction can enhance the detection of layered characteristics. Additionally, the results are generated almost in real-time, which facilitates quality assessment in the process of fabrication.
Figure 7 demonstrates the results of the grid-based fidelity analysis. Binarized reference templates Tn* and binarized foreground images Fn* for the same layers are combined, with Tn* represented in the green channel and Fn* in the magenta channel, as depicted in Fig. 7(a). The grid size for analysis is set at 20 × 20 pixels. In Fig. 7(b) and Fig. 7(c), we illustrate two cases of grid analysis conducted on layer 5 and layer 50. Cyan grids indicate regions where the fidelity value q exceeds 70%, whereas red grids highlight regions with a lower q and potential defects. From our observation, errors are more prevalent in grids with adjacent fibers where proximity effects are more likely to occur. However, the overall high fidelity shows that the fabricated structure successfully reproduces the original 3D micropatterns. In Fig. 7(d), we map the probability of error occurrence across all layers for each grid. Notably, one grid in the bottom left corner shows a significantly higher chance of errors at 23.81%, which could be attributed to the reduced laser power at the edge of the printing FOV.
Fig. 7. Grid-based fidelity analysis for the TPP-fabricated tissue scaffold. (a) Two-color overlap of Tn* and Fn*. Flaw detection at (b) z = 5 µm and (c) z = 50 µm. (d) Error probability across segmented grids. (e) q values for each layer over the entire z stack.
As shown in Fig. 7(e), the average q across the entire set of 60 layers is calculated to be 87.5%, with layer 51 exhibiting the lowest value. We observe a decreasing trend in q along with the increase of z. Additionally, from Fn* at higher layers, we find that there are more residues from the previous layers that cannot be eliminated. These artifacts impact the calculation of fidelity as they originate from detection rather than the printing process itself. This issue is mitigated by using Tn* to categorize these residues in dark regions. However, it imposes limitations on the maximum height that this monitoring method can achieve. Another aspect to consider is the axial resolution, which could potentially be enhanced by employing a light source with a shorter wavelength and narrower bandwidth. It is important to note that our method is tailored for layerwise fabrication and does not extend to other fabrication schemes such as random access scanning which is more efficient for creating complex hollow structures [23]. However, given the layerwise scheme is used routinely in the field, our method remains highly applicable to most systems.
4. Conclusion
In this study, we presented ABLE as a new approach for in situ monitoring of embedded 3D micropatterns by TPP-DLW. We estimated the system’s impulse response to identify optimal processing parameters and to create reference templates for comparative analysis. Adaptive background subtraction and grid-based analysis were implemented in real-time for assessing part quality. As a practical application, we demonstrated the inspection of a fibrous tissue scaffold derived from SHG data of bovine tendons, exemplifying ABLE’s effectiveness in analyzing internally complex structures. We obtained system-adjusted fidelity measures in situ with an average value of 87.5% and identified flaws in the TPP-fabricated structure.
This approach outperforms previous monitoring methods in terms of speed and automaticity. It only adds 34 ms to data acquisition time per layer and requires no hardware modifications. It will be a powerful add-on tool to improve TPP fabrication quality, and also pave the way for remote inspection and collaboration of TPP fabrication [24,25] as it requires much less storage than conventional video recording of the processes. Moreover, our technique would be particularly useful for TPP biomanufacturing where scaffolds with complex internal patterns are challenging and costly to characterize and inspect.
Appendix
For molecular systems absorbing two photons with the same frequency, the total number of molecules nabs excited by TPA per unit volume per unit time nabs is given by [26]
Funding
National Science Foundation (CMMI-2043168, CMMI-2043243).
Acknowledgments
We thank Dr. Christopher Rowlands for assistance in building the galvo scanning system. We also thank Zichang Gao for algorithm implementation. J.A.B. thanks Hibbitt Post-Doctoral Fellowship in the Brown School of Engineering and the Postdoctoral Diversity Enrichment Program of the Burroughs Welcome Fund.
Disclosures
The authors declare no conflicts of interest.
Data availability
Data underlying the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request.
References
1. S. Maruo, O. Nakamura, and S. Kawata, “Three-dimensional microfabrication with two-photon-absorbed photopolymerization,” Opt. Lett. 22(2), 132 (1997). [CrossRef]
2. M. McLamb, Y. Li, P. Stinson, et al., “Metasurfaces for the infrared spectral range fabricated using two-photon polymerization,” Thin Solid Films 721, 138548 (2021). [CrossRef]
3. H. J. McLennan, A. J. Blanch, S. J. Wallace, et al., “Nano-liter perfusion microfluidic device made entirely by two-photon polymerization for dynamic cell culture with easy cell recovery,” Sci. Rep. 13(1), 562 (2023). [CrossRef]
4. S. Thiele, K. Arzenbacher, T. Gissibl, et al., “3D-printed eagle eye: Compound microlens system for foveated imaging,” Sci. Adv. 3(2), e1602655 (2017). [CrossRef]
5. T. WeiÃ, G. Hildebrand, R. Schade, et al., “Two-Photon polymerization for microfabrication of three-dimensional scaffolds for tissue engineering application,” Eng. Life Sci. 9(5), 384 (2009). [CrossRef]
6. V. Ajeti, C.-H. Lien, S.-J. Chen, et al., “Image-inspired 3D multiphoton excited fabrication of extracellular matrix structures by modulated raster scanning,” Opt. Express 21(21), 25346 (2013). [CrossRef]
7. J. C. Culver, J. C. Hoffmann, R. A. Poché, et al., “Three-dimensional biomimetic patterning in hydrogels to guide cellular organization,” Adv. Mater. 24(17), 2344–2348 (2012). [CrossRef]
8. L. Muntean, R. Planques, A. L. D. Kilcoyne, et al., “Chemical mapping of polymer photoresists by scanning transmission x-ray microscopy,” J. Vac. Sci. Technol. B Microelectron. Nanometer Struct. Process. Meas. Phenom. 23(4), 1630 (2005). [CrossRef]
9. H.-B. Sun, T. Tanaka, K. Takada, et al., “Two-photon photopolymerization and diagnosis of three-dimensional microstructures containing fluorescent dyes,” Appl. Phys. Lett. 79(10), 1411–1413 (2001). [CrossRef]
10. C. Arnoux, L. A. Pérez-Covarrubias, A. Khaldi, et al., “Understanding and overcoming proximity effects in multi-spot two-photon direct laser writing,” Addit. Manuf. 49, 102491 (2022). [CrossRef]
11. S. K. Saha, C. Divin, J. A. Cuadra, et al., “Effect of Proximity of Features on the Damage Threshold During Submicron Additive Manufacturing Via Two-Photon Polymerization,” J. Micro Nano-Manuf. 5(3), 031002 (2017). [CrossRef]
12. R. Zvagelsky, F. Mayer, D. Beutel, et al., “Towards in-situ diagnostics of multi-photon 3D laser printing using optical coherence tomography,” Light Adv. Manuf. 3(3), 1 (2022). [CrossRef]
13. Y. He, Q. Shao, S. C. Chen, et al., “Characterization of two-photon photopolymerization fabrication using high-speed optical diffraction tomography,” Addit. Manuf. 60, 103293 (2022). [CrossRef]
14. T. Baldacchini, R. Zadoyan, M. Wegener, et al., “In situ and real time monitoring of two-photon polymerization using broadband coherent anti-Stokes Raman scattering microscopy,” Opt. Express 18(18), 19219 (2010). [CrossRef]
15. L. Kallioniemi, S. Annurakshita, and G. Bautista, “Third-harmonic generation microscopy of undeveloped photopolymerized structures,” OSA Contin. 3(11), 2961 (2020). [CrossRef]
16. E. Yulianto, S. Chatterjee, V. Purlys, et al., “Imaging of latent three-dimensional exposure patterns created by direct laser writing in photoresists,” Appl. Surf. Sci. 479, 822–827 (2019). [CrossRef]
17. X. Y. Lee, S. K. Saha, S. Sarkar, et al., “Automated detection of part quality during two-photon lithography via deep learning,” Addit. Manuf. 36, 101444 (2020). [CrossRef]
18. M. Abdelrahman and T. L. Starr, “Quality certification and control of polymer laser sintering: layerwise temperature monitoring using thermal imaging,” Int. J. Adv. Manuf. Technol. 84(5), 831–842 (2015). [CrossRef]
19. F. R. Velasco, “Thresholding using the ISODATA clustering algorithm,” IEEE Trans. Syst. Man Cybern. 10(11), 771–774 (1980). [CrossRef]
20. M. A. Fischler and R. C. Bolles, “Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography,” Commun. ACM 24(6), 381–395 (1981). [CrossRef]
21. D. G. Lowe, “Distinctive Image Features from Scale-Invariant Keypoints,” Int. J. Comput. Vis. 60(2), 91–110 (2004). [CrossRef]
22. C. Steger, “An unbiased detector of curvilinear structures,” IEEE Trans. Pattern Anal. Mach. Intell. 20(2), 113–125 (1998). [CrossRef]
23. Q. Geng, D. Wang, P. Chen, et al., “Ultrafast multi-focus 3-D nano-fabrication based on two-photon polymerization,” Nat. Commun. 10(1), 2179 (2019). [CrossRef]
24. Y. Yang, V. A. Kelkar, H. S. Rajput, et al., “Machine-learning-enabled geometric compliance improvement in two-photon lithography without hardware modifications,” J. Manuf. Process. 76, 841–849 (2022). [CrossRef]
25. S. Jia, J. Sun, A. Howes, et al., “Hybrid physics-guided data-driven modeling for generalizable geometric accuracy prediction and improvement in two-photon lithography,” J. Manuf. Process. 110, 202–210 (2024). [CrossRef]
26. M. Rumi, S. Barlow, J. Wang, et al., “Two-Photon Absorbing Materials and Two-Photon-Induced Chemistry,” Adv. Polym. Sci. 213(1), 1–95 (2008). [CrossRef]
