2.1.

Cell Specimen Preparation

HeLa cells (CVCL_0030v, ATCC) were cultured at 37°C, 5% ${\mathrm{CO}}_2$, in Dulbecco’s modified Eagle medium (DMEM) (11960044, Thermo Fisher, Waltham, Massachusetts, United States) containing 1% penicillin–streptomycin (P4458-100ML, Merck, Rahway, New Jersey, United States), 1% l-glutamine (392-0441, VWR), and 10% fetal bovine serum (10500064, Gibco, Waltham, Massachusetts, United States). Coverslips were coated with a 1:500 dilution of fibronectin bovine plasma (F1141-1MG, Merck) and incubated for 24 h prior to fixation.

Coverslips were fixed with 4% paraformaldehyde (PFA) (16005, Merck) at 37°C for 15 min, before being blocked and permeabilized for 30 min at 37°C in an immunofluorescence (IF) buffer comprising 2.5% goat serum (ab7481, Abcam, Cambridge, United Kingdom) and 0.3% TritonX-100 (X100, Merck) in phosphate-buffered saline (PBS), labeled using anti-tubulin rat monoclonal conjugated with Alexa Fluor 488 (ab195883, Abcam) at a dilution of 1:100 in IF buffer and incubated at 4°C for 24 h in a dark culture dish with stable humidity. The samples were then mounted using ProLong Glass (P36980, Thermo Fisher) to reduce photobleaching during prolonged imaging and left to set for at least 24 h prior to imaging.

For non-filamentous structures, a stable cell line of HA-GLUT4-GFP expressing 3T3-L1 cells was previously generated in the lab²⁵ plated on 18-mm-diameter coverslips and grown in DMEM (52100047, Gibco) supplemented with 10% new-born calf serum (26010066, Gibco), 5% l-glutamine, 5% penicillin–streptomycin and maintained at 37°C, and 10% ${\mathrm{CO}}_2$. Following differentiation into adipocyte²⁶ cells were put on ice, washed once with ice-cold 1X PBS, and fixed in 4% PFA for 30 min at room temperature. Coverslips were washed three times with 1X PBS before mounting on a slide with ProLong Glass. Specimens were allowed to set fully prior to imaging.

2.2.

Mesolens Imaging and Data Acquisition

Diffraction-limited images were obtained with the Mesolens using the widefield fluorescence illumination modality, a diagram of this system setup is shown in Fig. 2. Specimens were illuminated using a $490\pm 20\mathrm{nm}$ light-emitting diode (LED) sourced from a multi-wavelength illuminator (pE-4000, CoolLED, Andover, United Kingdom), and fluorescence emission was detected at $525\pm 25\mathrm{nm}$. The maximum power of the excitation source at the specimen plane was $35.7\pm 0.1\mathrm{mW}$. Imaging was carried out with the Mesolens correction collars set for water immersion to minimize spherical aberration. To ensure statistically independent images for SRRF processing, a time-lapse series of the fixed samples was taken to obtain 40 images, with the interval between image capture typically set to 2 to 3 s as the samples are fixed. Other works utilizing SRRF can use up to hundreds of individual images²⁷ to produce the final reconstruction; however, due to the size of the Mesolens image datasets and the computational restrictions this entails the datasets here are limited to 40 images for SRRF processing. This restricted dataset may reduce the ability of SRRF processing to suppress noise to avoid this exposure time was typically kept above 1500 ms when imaging, as longer exposure times would result in a higher SNR.

Fig. 2

Schematic diagram of the widefield epifluorescence Mesolens system.

To ensure the diffraction-limited widefield Mesolens images captured are of the highest resolution, images are captured using a camera with a chip-shifting sensor (VNP-29MC, Vieworks, Anyang-si, South Korea). This shifting moves the 29-megapixel chip in a $3\times 3$ array, giving an image resolution of $\mathrm{19,728}\times \mathrm{13,152}\text{pixels}$ for a total 259.5-megapixel capture. Over the $4.4\times 3.0\mathrm{mm}$ FOV, this results in a 226-nm pixel size—satisfying Nyquist sampling.²⁸ The acquisition of one full FOV image using the sensor shifting camera took 13.5 s when using a 1500-ms exposure. When including the additional time for the transfer of image data to the computer, and the interval between captures, imaging of a 40-image stack typically took $\sim 22\min$.

2.3.

SRRF Processing

The diffraction-limited widefield Mesolens images were opened within Fiji.²⁹ To avoid the possibility of introducing any artifacts no preliminary processing was carried out, instead using the Nano-J SRRF plugin¹⁶ the raw images can be processed. There were restrictions on the parameters that can be used when processing Mesolens images due to the unconventional file size of the datasets, and as such the parameters used vary slightly from the default values. The ring radius was set to 1.90—when carrying out quantification of the image resolution this was found to result in the highest resolution while also allowing for ease of processing. The radiality magnification was reduced so that the large datasets could be processed. This was kept at two as it allowed for SRRF processing of intact Mesolens datasets. If the magnification was increased further processing could only be carried out on small ROIs, which would introduce stitching and tiling artifacts. For the temporal analysis, temporal radiality pairwise product mean (TRPPM) was selected as the fluorophores used do not blink but instead have slight intensity fluctuations which TRPPM is best suited to process. Using TRPPM over alternative options such as temporal radiality average is also beneficial as it provides additional noise suppression, which is beneficial given the limitations on the size of the datasets used.¹⁵ Any unmentioned parameters remained at the default value. The typical processing time of one dataset was $\sim 95\min$, including opening the dataset within Fiji. The computational power available was limited, and this figure could be greatly reduced in further iterations.

2.4.

SQUIRREL Analysis and Resolution Measurements

Following SRRF processing, the accuracy of the output image was first assessed using SQUIRREL. Due to the size of the processed SRRF images, SQUIRREL analysis could not be conducted on the intact image. As there is no radiality component within SQUIRREL, the image was split into tiles and later recombined following processing without the introduction of artifacts on the edge of each tile. Macros were used within Fiji to split both the reference image and super-resolution SRRF image into tiles of equal sizes, with the required diffraction-limited reference image¹⁹ taken as the first diffraction-limited widefield Mesolens image from each stack. From here, it was possible to perform SQUIRREL analysis on each of the matching tile pairs, with the RSF estimated through optimization each time, and the individual SQUIRREL error tiles could then be stitched³⁰ into one complete error map. The individual RSP and RSE values calculated for each tile were averaged to determine the value for the complete image, with the error taken as the standard deviation. RSP measures the Pearson correlation coefficient among the images on a normalized scale between $-1$ and 1, with one being identical images. This acts as a measurement of the structural correlation among images and is independent of intensity. RSE is an intensity-dependent measurement of the root-mean-square error among images, where lower values indicate better agreement between the raw image and the super-resolution equivalent. For ease of comparison among datasets, the RSE value was normalized against the total range of possible values. Typically, each Mesolens image is split into 36 tiles for processing, and SQUIRREL analysis of each pair of tiles takes $\sim 1\mathrm{h}$.

The resolution was measured using image decorrelation analysis²⁰ Fiji plugin. Due to restrictions imposed by the large Mesolens file size, the measurements were carried out on $n=3$ ROIs from across the field of each super-resolution reconstruction, and the final resolution for each image is an average of these ROIs. Each ROI was cropped from the original super-resolution SRRF Mesolens images, with no additional image processing prior to this stage to ensure an accurate measurement. The ROIs were opened within Fiji and then processed using image decorrelation analysis and the default settings. The resultant plots show the decorrelation curves of the ROI, along with the calculated cutoff frequency, $k_c$, and the amplitude of the local maximum of the decorrelation function, $A_0$. The achievable resolution presented here is an average of $N=9$ ROIs, across $N=3$ tubulin-labeled datasets—again the error is taken as the standard deviation.

3.1.

Applications to Filamentous Structures

Figure 3(a) shows a full Mesolens FOV image of tubulin-labeled HeLa cells, following SRRF processing, demonstrating the scale and the volume of cells captured per image, along with the absence of any stitching and tiling artifacts across the entire FOV of super-resolved reconstruction. Digital zooms of ROI are highlighted to demonstrate the improvement in contrast and resolution between the original Mesolens images [Figs. 3(b)–3(d)] and the corresponding SRRF equivalent [Figs. 3(e)–3(g)]. Between each of the corresponding ROI pairs, there is a clear improvement in image clarity and spatial resolution, allowing for a level of detail not previously visualized across the mesoscale. No additional image processing was carried out on either set of images.

Fig. 3

Complete Mesolens FOV showing the SRRF-processed image of the tubulin in HeLa cells labeled with AF488 with three highlighted ROI: magenta, cyan, and yellow (a). A digital zoom of the raw widefield epifluorescence diffraction limited Mesolens image is shown for each ROI (b)–(d), with a digital zoom of the improved ROI following SRRF processing (e)–(g). Scale bars are $30\mu \mathrm{m}$ for all ROIs.

SQUIRREL analysis on the dataset shown in Fig. 3 is presented in Fig. 4. These data show there is minimal error and high levels of agreement between the original and the super-resolution images across the full FOV [Fig. 4(a)]. The same ROIs in Fig. 3 are shown with SQUIRREL processing in Figs. 4(b)–4(d). These show the areas of disagreement between the raw Mesolens image and the super-resolution equivalent. These ROIs highlight that this low error and high agreement is continuous across both the full FOV and the individual regions and that although there are areas of disagreement across the map, the disagreement is small. Figures 4(e)–4(g) show digital zooms of a contrast-adjusted error map. The image was adjusted using contrast-limited adaptive histogram equalization (CLAHE) with the default parameters (block size = 127, histogram bins = 256, maximum slope= 3.0, mask = none, and fast = false)³¹ to highlight the areas of discrepancy. We note these discrepancies are largely confined to areas where the image is saturated.

Fig. 4

Complete reconstructed SQUIRREL error map showing the areas of agreement between the raw widefield epifluorescence diffraction limited Mesolens image and the SRRF reconstruction (a). In the middle row, a digital zoom of the error map is displayed for each ROI (b)–(d), and in the bottom row, a contrast-adjusted ROI highlighting the areas of error (e)–(g). Scale bars are $30\mu \mathrm{m}$ for all ROIs.

For the dataset shown in Figs. 3 and 4, this high fidelity among images is confirmed by the RSP value of $0.989\pm 0.004$. RSP is a normalized value between $-1$ and 1, with one representing structurally identical images. The value shown here indicates high structural agreement, and minimal artifacts between both images, alongside the intensity-dependent RSE result of $1.13\%\pm 0.23\%$. The RSE value presented has been normalized against the total intensity range in the SQUIRREL error map, allowing for simplified comparisons among different datasets. The resolution of this super-resolution reconstruction in Fig. 3(a) was also measured from three ROI at 442.9 nm [see Figs. S1(a)–1(c) in the Supplementary Material]. A significant improvement in image resolution surpasses the theoretical maximum resolution achievable on the Mesolens—providing super-resolved images.

To calculate an average resolution measurement achieved using this method, measurements were repeated for three biological repeats of HeLa cells where tubulin was labeled with an AF488 antibody conjugate. There was an average RSP value of $0.991\pm 0.006$ and an average RSE value of $0.81\%\pm 0.30\%$, with consistently low error rates and high agreement across different images. An average resolution was also recorded using three ROI from each sample and measured at $446.3\pm 10.9\mathrm{nm}$ (see Fig. S1 in the Supplementary Material).

Due to the size of the Mesolens data, it is currently not feasible to increase the SRRF magnification above 2 on raw data without surpassing the maximum allowed file size for processing. For higher magnification values, SRRF processing is possible on ROIs only, so a magnification of 2 is used here to allow for the processing of raw Mesolens data without the addition of stitching and tiling artifacts. To demonstrate that using a higher magnification would have minimal impact on image resolution, SRRF was carried out on the same ROIs as shown in Fig. 3, using increasing magnification and the average resolution across these ROIs for each magnification was calculated, as shown in Fig. 5.

Fig. 5

Image resolution measured using image decorrelation analysis, for increasing SRRF magnification, comparative to the previously shown Mesolens resolution of 700 nm.²³ Error bars represent the standard deviation.

3.2.

Applications to Non-Filamentous Structures

To demonstrate this method is not confined to filamentous structures, Fig. 6 shows a full Mesolens FOV capture of green fluorescent protein (GFP) tagged GLUT4 glucose transporters in 3T3-L1 fibroblasts following SRRF processing, with digitally magnified ROIs to again show the improvement in the SRRF reconstruction from the original widefield epifluorescence diffraction limited Mesolens image. Figure 6(a) shows the entire FOV of super-resolved reconstruction, and the intact cells captured again without any stitching and tiling artifacts. Digital zooms of ROIs demonstrate the improvement in contrast and resolution between the original Mesolens images [Figs. 6(b)–6(d)] and SRRF equivalent [Figs. 6(e)–6(g)]. Again, between each of the ROI pairs, there is a clear improvement in resolution and image clarity, allowing for clearer visualization of the GLUT4 molecules throughout the cell interior. GLUT4 is sequestered within multiple different intracellular compartments in the absence of insulin,^25,32 and the images shown here provide a clear illustration of this distribution. The resolution of this image was measured from three ROIs at $431.6\pm 9.8\mathrm{nm}$ (see Fig. S2 in the Supplementary Material), which again surpasses the maximum optical resolution theoretically possible on the Mesolens, obtaining super-resolved images.

Fig. 6

Complete Mesolens FOV showing the SRRF processed image of the GLUT-4 in 3T3-L1 adipocytes expressing GFP, with three highlighted ROI: magenta, cyan, and yellow (a). A digital zoom of the raw widefield epifluorescence diffraction limited Mesolens image is shown for each ROI (b)–(d), with a digital zoom of the improved ROI following SRRF processing below (e)–(g). Scale bars are $30\mu \mathrm{m}$ for all ROIs.

SQUIRREL analysis on this dataset is shown in Fig. 7. There is again minimal error and high levels of agreement between the raw widefield epifluorescence diffraction limited Mesolens image and the SRRF reconstruction throughout across the full FOV [Fig. 7(a)]. Digital zooms of the ROIs [Figs. 7(b)–7(d)] highlight that there is some error present across the FOV, but it is minimal. Figures 7(e)–7(g) show digital zooms of a contrast-adjusted error map. The image was adjusted using CLAHE with the default parameters (block size = 127, histogram bins = 256, maximum slope = 3.0, mask = none, and fast = false)³¹ to better highlight the areas of discrepancy. For this dataset, the average RSP value was calculated as $0.993\pm 0.003$ and the RSE value as $0.68\%\pm 0.14\%$, again demonstrating the high fidelity and low levels of error between the original raw diffraction limited widefield Mesolens image and the SRRF processed reconstruction.

Fig. 7

Reconstructed SQUIRREL error map showing the areas of agreement between the raw widefield epifluorescence diffraction limited Mesolens image and the SRRF reconstruction with highlighted ROI: magenta, cyan, and yellow (a). A digital zoom of the error map is displayed for each ROI (b)–(d), with a contrast-adjusted ROI highlighting the areas of error within each ROI below (e)–(g). Scale bars are $30\mu \mathrm{m}$ for each ROI.