2.1.

Optical Setup

Fluorescence imaging was performed using a light sheet microscope (Fig. 1), whose optical system is substantially similar to the ultramicroscope developed by Dodt.¹¹ To excite the autofluorescence, a diode-pumped solid-state laser emitting at 473 nm (CNI, MBL-FN-473) is coupled into a single-mode optical fiber that delivers the light to the setup. The laser beam is split into two portions in order to illuminate the sample from opposite sides: two cylindrical lenses ($f=75\mathrm{mm}$) are used to create a light sheet made by two beams counter-propagating across the sample. The illumination power was 5 mW for each side. A chopper wheel in the illumination path allows one to illuminate the sample alternatively from opposing directions in order to reduce the effect of light sheet spreading caused by light scattering.¹⁵ Rotation of the sample around 360 deg is also possible to support the multiview approach.¹⁶ A $4\times$ objective (Nikon, PLAN FLUOR, $\mathrm{NA}=0.13$) with a long working distance (17.2 mm) is used to collect the fluorescence emitted orthogonally to the light sheet. The objective, in combination with a tube lens (Nikon) and a long pass filter (FELH500, Thorlabs), forms the image of the fluorescence emission on a 4 MP CMOS camera (ORCA Flash 4.0, Hamamatsu, $n_x=2048$ and $n_y=2048$), with a field of view of ca. $3\times 3{\mathrm{mm}}^2$. The lateral resolution was limited by the $\mathrm{NA}=0.13$ objective to about $2.3\mu \mathrm{m}$ and the later resolution was limited by the waist of the light sheet, which was $15\mu \mathrm{m}$. The camera is synchronized with the chopper in order to alternatively acquire the fluorescence images corresponding to the two counter-propagating light sheets. Each pair of acquired images, corresponding to illumination from the two opposite sides, was computationally fused to yield superior image quality.¹⁵ In particular, the halves of the images that present more artifacts (which are due to absorption and scattering of the excitation beams) were removed and the remaining two halves were merged using a stitching algorithm.¹⁷ Exposure times were 100 ms for each plane, depending on the fluorescence signal. The typical number of acquired planes was $n_z=400$, with steps of $10\mu \mathrm{m}$ and the total acquisition time was 40 s.

Fig. 1

Sketch of the optical setup. The sample is immersed in a glass cuvette filled with clearing liquid and is scanned though the detection image plane by a motorized stage (not shown). The displayed elements include the single-mode (SM) optical fiber, mirrors (M), achromatic cylindrical lens (ACL) and the CMOS camera.

2.2.

Animals

Six, 8-week-old, Balb C mice were prepared for the experiment using enforced gavage with water solution for three consecutive days. Then the mice were euthanized with an overdose of Avertine (tribromoethanol, $250\mathrm{mg}/\mathrm{kg}$; Sigma Aldrich, T48402) and the whole intestine was collected from each animal for analysis. All the procedures with mice were approved by the local Ethics Committee for Animal Welfare (IACUC N° 583) and were carried out in strict respect for European and National regulations.

2.3.

Sample Preparation and Measurement Procedure

After excision, the intestines underwent fixation with formalin (Sigma Aldrich, F8775) at 4°C for 24 h, dehydration in methanol (33%, 66%, 100%, 15 min/step), overnight bleaching with freshly prepared $\mathrm{MeOH}:{\mathrm{H}}_2{\mathrm{O}}_2:\mathrm{DMSO}$ buffer in a 2∶1∶3 ratio, and chemical clearing in a 1∶2 solution of benzyl alcohol-benzyl benzoate (BABB) to reduce optical scattering. The samples were then cut into chunks approximately 10 mm long. For the LSFM experiment, each piece was held in place in a cuvette filled with BABB using small tweezers and moved through the light sheet, using a motorized translation stage (PI, M-405.CG) to obtain a 3-D image stack.

2.4.

Virtual Unfolding

The application of LSFM to each intestine sample led to an image stack of about $3\times 3\times 3{\mathrm{mm}}^3$ volumes with $n_x·n_y·n_z$ voxels. The dataset can be visualized in Cartesian coordinates along transverse, sagittal, and coronal sections (Figs. 2 and 3) or in 3-D reconstructions. Nevertheless, such views do not match the geometry of the sample, which could be better analyzed in a reference system based on the external shape. More specifically, the sample could be conveniently visualized layer by layer, from the outward surface to the inner bore, or vice versa. For this purpose, we processed the acquired dataset using an automatic routine that projects the sample in an appropriate framework (Fig. 4). The procedure consists of

Fig. 2

Sections of the sample. (a) Coronal section, (b) sagittal section, and (c) transverse section. Scale bar $500\mu \mathrm{m}$ (Video 1, 6.98 MB) [URL: http://dx.doi.org/10.1117/1.JBO.21.5.056001.1]..

Fig. 3

Coronal and transverse sections of an untreated intestine in different anatomical regions. Scale bar $500\mu \mathrm{m}$.

Fig. 4

Upper panels: scheme of the procedure: a sample with cylinder-like shape is “unfolded” and projected to a planar geometry. Lower panel: scheme of the procedure that allows one to select $N$ points that lay on each of the $M$ layers of the tissue. The procedure is repeated in all the transverse sections of the sample. Scale bar $500\mu \mathrm{m}$.

The method creates $M\times N$ new point-like samples in each of the $n_y$ transverse sections, leading to a new image stack of $M\times N\times n_y$ voxels that contains only the relevant information, because the majority of the meaningless voxels (black voxels) inside and outside the intestine are removed. This usually brings a reduction of about twofold of the dataset. More importantly, this new dataset results from the projection of the original one onto a reference system based on the shape of the sample, which is more suitable for the analysis of the tissue morphology. In particular, the 2-D images of $M$ tissue layers, spaced one another by a quantity $\mathrm{\Delta }r$, are intrinsically embedded in the dataset and can be easily displayed and processed.

The processing is run automatically once the user selects a threshold for segmentation, the sampling distance $\mathrm{\Delta }{s}_0$, and the total distance $M·\mathrm{\Delta }r$. The distance $\mathrm{\Delta }{s}_0$ indicates the spatial sampling on the surface of the sample. The distance $M·\mathrm{\Delta }r$, that determines the inner position in which the processing is stopped was normally $300\mu \mathrm{m}$ for the analyzed intestines (with $M=100$). The value $N$ was determined by the sampling distance $\mathrm{\Delta }{s}_0$, which normally is equal to or is a multiple of the pixel size. Depending on the sample size, $N$ was between 3000 and 6000. Since the perimeter of the sample decreases during the erosion, $\mathrm{\Delta }s$ decreases as well. This means that different layers are characterized by different sizes. For this reason, each layer is stretched (with a linear interpolation) by a factor $\mathrm{\Delta }s/\mathrm{\Delta }{s}_0$, creating a stack of layers that are aligned.

The data processing required ca. 4 min per stack with a i7 4930 CPU, 3.4 GHz. The MATLAB code for virtual unfolding is available upon request.

2.5.

Segmentation of the Villi

Villi segmentation was performed on the $M$ images of the intestine layers obtained after the virtual unfolding procedure. The so called “watershed segmentation” method,¹⁸ consisting of a series of grayscale morphological image processing operations was implemented in MATLAB. Each villus was automatically colored to distinguish it from adjacent ones. The code for segmentation is available upon request.

3.1.

Reconstruction of Large Portions of the Intestine with Light Sheet Fluorescence Microscopy

We observed strong and similar autofluorescence intensities in the mucosa and muscular tissues of different intestine samples. Biological tissues exhibit a noticeable autofluorescence that can be ascribed to several fluorophores. Among them, we can mention amino acids, structural proteins, notably collagen and elastin, the coenzyme nicotinamide adenine dinucleotide in the reduced form (NADH), and flavin derivatives, like riboflavin.^19,20 An increase in tissue autofluorescence is also rather common after fixation with formalin due to cross-link of dopamine.^21,22 In the specific case of the intestine, considering the excitation (473 nm) and the emission ($>500\mathrm{nm}$) wavelengths, the most likely candidates for the autofluorescence observed in the present experiment are flavin derivatives, like the flavin mono (FMN) and dinucleotide (FAD), with possible contributions from lipopigments, like lipofuscin, and enteramine.²³ All these fluorescent compounds have been documented to contribute to the intestine autofluorescence upon excitation with two photons, roughly corresponding to blue light.²⁴ As a matter of fact, the fluorescence observed in a thin section of fixed intestine is strong, as shown in Fig. 5. In order to assess the morphology of the intestine in various regions, we imaged the tissue autofluorescence and visualized the acquired volumes in different planes (sections) or in 3-D renderings (Fig. 2 and Video 1).

Fig. 5

Thin section of a fixed intestine sample. (a) H&E and (b) autofluorescence. Scale bars $100\mu \mathrm{m}$.

LSFM provided high-resolution volumetric reconstruction of the samples, showing fine anatomical details of villi, glands, and muscular tissue. In particular, the muscular tissue close to the outer surface of the sample was clearly recognizable by the typical circularly shaped muscular cells aggregates (Video 1). Remarkably, light-sheet shadowing effects caused by light attenuation, which occasionally downgrade the image quality in LSFM, were very limited over the entire intestine. This is likely due to the low density of the intestinal tissue compared to, e.g., brain or spinal cord. Moreover, the sample presents a cavity in its center, which contributes to reduce the attenuation of the excitation light. Indeed, the multiview acquisition was not necessary and high quality reconstructions of 2- to 3-mm-thick intestine tissues were obtained in a single scan. The high quality achievable with a single acquisition facilitates the analysis of a large number of samples, reducing the time required for alignment and image acquisition. Multiple view acquisition could be used to further improve the resolution or to achieve isotropic resolution, but at the cost of longer acquisition and processing time, which is not always compatible with the fact that one may need to analyze hundreds of samples in a clinical study.

Various regions of the intestine, representative of different anatomical districts (duodenum, ileum, colon), are shown in Fig. 3. We observe that the autofluorescence intensity decreases going from the upper tract of the intestine (duodenum) to the lower part (colon). This decrease was systematically observed in all the samples that were entirely analyzed ($n=6$). Conversely, the distribution and shape of the villi, the morphology of the wall, and of the internal structure of the organ are clearly distinctive along the different tracts.

3.2.

Layer by Layer Visualization

A detailed analysis of villus morphology, coupled to a quantitative spatial assessment of villi density along the intestine, is mandatory to either detect or monitor the extension of the putative intestinal lesions along its whole length.⁸ Although a large number of distinct villi could be imaged at high resolution in each acquired region, quantification of their density was computationally expensive and inaccurate within the 3-D volume. Using the standard imaging approach, a relatively large amount of data ($>4\mathrm{GB}$) was accumulated for each acquisition, leading to a strong computational effort for the segmentation of the different compartments and the localization of the villi. In particular, identification of villi was challenged by the fact that they could lay in any direction within the 3-D volume. Moreover, the majority of the stored data corresponded to black pixels belonging to the volume external to the tissue and to its internal cavity. We, therefore, looked for a more efficient way to visualize and process the image stacks.

For this purpose, we developed an algorithm to virtually unfold the intestine around its external surface by exploiting its cylinder-like shape. The algorithm projects the layers of the sample in a planar geometry (Fig. 4, upper panels). A similar method was proposed by Schmid et al.²⁵ for spherically shaped specimens. Yet, although the intestine presents a cylindrical symmetry in many tracts, this is not the case along its whole length. Hence, a simple extension of the Schmid method to the projection of the dataset into a cylindrical coordinate system could not be applied. As an alternative method, we tried to sample the inner layers of the tissue by selecting a sequence of equally spaced points along the set of lines normal to the boundary of the sample. Yet, this method was too sensitive to the unevenness of the outer surface of the intestine that led to unacceptable segmentation noise. This problem has been recently solved by generating a smooth model of the surface of interest by fitting the segmented volume or through triangulation.²⁶

Alternatively, we devised a virtual unfolding routine based on morphological operations that takes into account the shape of the intestine. Each layer was obtained by the erosion of the external boundary of the organ, as described in Sec. 2. This approach can be applied in principle to any geometry, because the layers can be obtained by applying the morphological erosion directly to the 3-D model. However, in order to process the dataset more rapidly, we applied the method to each transverse section of the acquired stack, taking advantage of the elongated shape of the intestine.

3.3.

Quantification of Villi Density, Shape and Morphology

Once the intestine was virtually unfolded, we could navigate through its volume, layer by layer (Fig. 6). Going inward from the outer boundary, the first layers of the intestine are characterized by the presence of the superficial vessels and the muscular tissue. Going deeper, a transition region, rich in muscular tissue and villi, is observed. Finally, the villi that protrude from the epithelial lining of the intestinal wall are reached and cut perpendicularly to their axis (Fig. 6, lower panel). In each 2-D unfolded image, hundreds of villi are sectioned, producing a significant assay of their morphology, density, and shape. The sampling of each layer is performed at a constant distance $\mathrm{\Delta }s$ along the perimeter of the transverse sections of the specimen. This is crucial to reduce the deformations that would have been caused by projecting the sample to a simple cylindrical geometry. Indeed, the majority of the villi are cut transversally by the layers (Fig. 4), providing an accurate reconstruction of their shape in the new coordinate system. However, because of the irregular morphology of the tissue, some villi are cut obliquely or unevenly by the layer, particularly in the highly curved region of the sample. This can cause deformation of their shape (e.g., they can appear elliptical), but overall their number and density can still be measured accurately. The quantification of the villi number and density was accomplished in the new coordinate system using standard watershed segmentation (see Sec. 2). In Fig. 7, each segmented villus is shown in a different color. The separation between adjacent villi and the sharpness of the image indicates that not only the density and distribution of the villi can be obtained, but also their position and shape can be statistically analyzed. Looking at more details in the image of a layer and at the segmentation of the villi (colored panels in Fig. 7), we observe that the image quality varies along the horizontal direction. In fact, because of the unfolding, the center of the image corresponds to the portion of the sample closer to the microscope objective, while the extremities correspond to the back side. Hence, it is not surprising that the center of the image shows a resolution higher than the left and right borders: this difference is due to the scattering of the fluorescence emission while propagating through the sample toward the objective. Here, multiview acquisition would have improved the image quality of the dorsal region. However, the resolution obtained with a fast single-view scan was sufficient for accurate segmentation of the villi in the entire image.

Fig. 6

Results of the unfolding procedure. Layers of the intestine are visualized at three different locations (depths) showing vessels and muscular structures as well as the mucosa with its villi. Scale bar $500\mu \mathrm{m}$.

Fig. 7

Results of the segmentation analysis. (a) Original layer obtained after unfolding at a depth of ca. $150\mu \mathrm{m}$ inside the tissue. (c) Segmented image. (b and d) details of the segmentation in two regions, corresponding to the dorsal and frontal part of the tissue. Scale bars $500\mu \mathrm{m}$ (left) and $100\mu \mathrm{m}$ (right).

It is worth noting that Fig. 7 shows the result of an automatic segmentation. Comparison with manual segmentation of the villi showed a difference lower than 5% across the entire image. Hence, taking advantage of the quality of the images achieved with the unfolding procedure and of the layer navigation method, we were able to segment and statistically compare the anatomical features of multiple samples.

In two out of six mice, we see clear differences in shape and density of villi (Fig. 8, left side), mainly in the duodenum tract of the intestine. In this region, the normal villi density is $\delta =76\pm 15{\mathrm{mm}}^{-2}$ ($n=6$), calculated as the number of segmented villi divided by the outer surface of the 3-D volume. Such a lesion is variable among mice as well as along the whole intestine. Some mice developed a segmental hyperplastic lesion, with conserved density of villi, but enlarged villous size (region inside the red rectangle in the central panel of Fig. 8), others showed severe lesions creating regions characterized by flattened villi and crypt hyperplasia with reduction in villi density with respect to the outer surface (region inside the green rectangle in the lower panel of Fig. 8).

Fig. 8

Representative intestines showing no damage (normal), mildly and severely damaged tissues. At $200\mu \mathrm{m}$ in depth from the outer surface of the intestine, the unfolded layers (right side) show hyperplastic regions (red) and atrophic regions (green). Scale bars $500\mu \mathrm{m}$.

These results indicate that the analysis of intestine lesions is feasible in terms of the density of villi, which can be measured with an automatic procedure. The proposed method, based on LSFM, has several advantages with respect to serial histological sectioning. Most of all, large portions of the intestine can be analyzed in a quantitative way with minimal preparation of the sample, no staining or labeling, and fast acquisition and processing of data.

Many intestinal diseases, such as coeliac disease or chronic inflammatory diseases, manifest a scattered lesion pattern that affects only parts of the entire organ, while others are preserved. In addition, intestinal lesions may be restricted to the mucosal surface or may also involve the whole mucosa within a selected region. The assessment of both severity and extension of intestinal lesions in experimental models of intestinal diseases, along with the putative improvement in mucosal architecture during treatment, must imply an accurate comparison with a reference “normal” counterpart. Thus, coupling wide-field and high-resolution imaging may represent a priceless tool to understand disease evolution as well to monitor the putative effects of treatments in animal models of disease.