Introduction

Osteoarthritis (OA) is the most common chronic joint disease, primarily affecting patients aged 55 and older, causing pain and disability, and impairing quality of life1. OA presents a growing burden on health services and societies worldwide2.

It is a complex disease involving several tissues, including hyaline cartilage (HC), subchondral bone (SCB), calcified cartilage (CC) located between HC and SCB, and synovial fluid3. The osteochondral interface, which comprises SCB and CC is named subchondral plate and is an active component of the disease contributing to structural progression in OA4,5. The crosstalk between chondrocytes and osteocytes plays a role in initiating OA through various signaling pathways6. In mature cartilage, chondrocytes are the only cells to maintain its homeostasis7 while osteocytes are the master regulators of bone remodeling8. In OA, chondrocyte clusters express a wide range of activation and abnormal differentiation markers9, chondrocyte hypertrophy-like changes occur in both early and late stages of the disease10. Alterations in osteocyte morphology and density during OA progression are thought to be related to structural alterations in the SCB11,12. Some authors have recommended to analyze CC separately from HC and SCB to deepen understanding of the OA process13.

Various imaging methods are available to study the cartilage-bone interface, including light microscopy, confocal scanning laser microscopy, and micro-computed tomography (micro-CT). Quantitative image analysis typically involves an image segmentation step to obtain binarized objects of interest. Simple global thresholding based on voxels intensity is commonly used to segment cell lacunae in cortical bone in micro-CT and synchrotron radiation micro-CT (SR micro-CT) images14, as well as in synchrotron radiation phase contrast micro-computed tomography (SR phase contrast micro-CT) images15,16. This method is simple and fast, however the segmentation quality may be limited by the uneven illumination of original images, partial volume effect, image noise and artifacts, which corrupt the grey level histogram. Hysteresis thresholding has been used to segment the cell lacunae (osteocytes and chondrocytes) in SR phase contrast micro-CT images16,17. While this method is well-suited for segmenting homogeneous objects with varying shapes, and guarantees to get connected regions, it has the disadvantage of incorporating unwanted neighbors with similar intensities. In the present study, contacting tissues such as CC and SCB exhibit similar grey levels, leading to misclassification using simple or hysteresis thresholding methods. The marker-controlled watershed (MCW) method has previously demonstrated convincing segmentation in 3D SR micro-CT images of bone vascularization18. This method initializes the segmentation from pre-identified markers to reduce unwanted seeds and separates an image into different regions by constructing watershed lines on a control surface image.

In the present study, our aim is to classify chondrocytes in CC and osteocytes in SCB, and to further quantify their morphological and topological changes in both control and OA groups using SR phase contrast micro-CT images of human knees at high spatial resolution (voxel size \(\:0.65\:{\upmu\:}\text{m}\)). Although the previously developed MCW algorithm was promising for some samples, it was not completely satisfactory. We observed that it could lead to over-segmentation of chondrocytes due to the incorrectly identified markers and poor edge detection in the control surface image, as discussed in more detail below. Therefore, we propose combining the MCW algorithm with a deep learning method (nnU-Net) to segment cell lacunae (both chondrocytes and osteocytes). The segmented cell lacunae are further divided into chondrocytes and osteocytes by using a manual mask to separate CC from SCB. Although these two regions present similar absorption contrast, they can be distinguished thanks to the phase contrast. The proposed segmentation method was validated on an experimental dataset using the Dice coefficient, sensitivity and specificity. Finally, we quantitatively measured several parameters: calcified cartilage and cell volume, their volume fraction, lacunar number density, anisotropy and structure model index of chondrocytes in CC and osteocytes in SCB. A t-test and the general linear mixed model were used to compare these parameters between the control and OA groups.

Methods

Sample preparation

Twenty-one left knees from human white donor cadavers were obtained from the Institute of Anatomy, Paris, France. The sample included 19 females and 2 males with an average age of 83.33 ± 11.19 (mean ± standard deviation) years old. Knees were classified based on posterior-anterior radiographs (Axiom Luminos, Siemens, Munich, Germany) scored using the Kellgren-Lawrence system19. Bone cores were extracted from the medial tibial plateau in different areas (central, external, posterior) using a 7 mm diameter trephine under constant water irrigation (BROT, Argenteuil, France). For quantitative analysis, fifteen bone cores were selected from the central zone, not covered by meniscus, and analyzed. Of these, ten of them were considered as controls with a Kellgren-Lawrence (KL) score 0 or 1, while five as osteoarthritic with a KL score larger than 2. The samples were completely defatted prior to acquisition using supercritical CO220. Details of the experimental data can be found in Table S1 in the supplementary material.

No additional information was available regarding the cause of death, prior illnesses, or medical treatments for these subjects, except for the absence of hepatitis and human immunodeficiency virus. All subjects had willed their bodies to science in accordance with the Declaration of Helsinki and remained anonymous. All methods were carried out in accordance with relevant guidelines and regulations. The study was approved by our institutional review board and conducted according to protocols established by the Human Ethics Committee of the Institut d’Anatomie, Université Paris Cité. The use of human specimens was authorized by the French Ministry of Higher Education and Research (CODECOH number DC-2019-3422).

Synchrotron radiation phase contrast micro-computed tomography (SR phase contrast micro-CT)

Samples were imaged on the ANATOMIX beamline at the Synchrotron SOLEIL, Orsay, France21. A parallel white beam tuned to 40 keV was used. The detector was a CMOS ORCAA coupled with a LuAg scintillator. Magnification was set at \(\:\times\:10\) producing \(\:2048\times\:2048\) projection images with a voxel size of 0.65 μm. A total of 4000 projections were acquired over 360°, with an exposure time of 66.6 ms, and the total scan duration was approximatively 5 min. SR phase contrast micro-CT images were reconstructed with Paganin phase retrieval and filtered back projection with the PyHST2 software developed at ESRF22. The 3D reconstructed volumes consisted of \(\:2048\times\:2048\times\:2048\) voxels, with each a voxel size of 0.65 μm. An example of an SR Phase contrast micro-CT slice is shown in Fig. 1, compared to a histological image.

Fig. 1
figure 1

(A) Histological image of the osteochondral junction stained with Stevenel blue and picrofuchsine. (B) SR phase contrast micro-CT image averaged from 20 slices (equivalent thickness: 13 μm). HC: hyaline cartilage, CC: calcified cartilage, SCB: subchondral bone.

Full size image

Segmentation

The goal of the segmentation step in this study was to classify all voxels into four classes: cell lacunae (chondrocytes and osteocytes), canals, subchondral plate (SCP), and background. This task, however, is challenging due to voxel intensity (grey level) overlaps occurring in the different classes, leading to poor segmentation results using simple threshold-based methods (Fig. S1, Supplementary material).

Here, the general scheme of the method combining a watershed-based algorithm (MCW) with a nnU-Net deep learning method is outlined in Fig. 2. More implementation details can be found in the following subsections.

Fig. 2
figure 2

Simultaneous segmentation of cell lacunae (chondrocytes and osteocytes), canals, subchondral plate (calcified cartilage and subchondral bone), and background using the proposed protocol.

Full size image

MCW algorithm and connected component analysis

A MCW algorithm and connected component analysis (CCA) were used to generate a preliminary segmentation. The MCW algorithm was implemented in a shell script with the Insight Toolkit (ITK), an open source library for image analysis23,24.

The key step in the MCW was the creation of a control surface image, which allowed to generate a marker to initialize the watershed process. Three marker classes: subchondral plate (CC + SCB), background, and lacunae plus canals were automatically generated using hysteresis thresholding and morphological filters18. After rescaling the reconstructed images between 0 and 255, a hysteresis thresholding was applied to generate the subchondral plate marker using a high threshold of 200 and a low threshold of 180 in 8-bit images. For the lacunae and canals marker, we inverted the image and applied hysteresis thresholding with parameters of 240 and 230. A simple global threshold was found sufficient to extract the background marker, based on its uniform intensity of 128 grey level in the images. We selected the parameters according to the grey level distributions in the images. Since the control and OA samples exhibited similar intensity distributions, the same parameters were applied across all samples. In addition, a morphological erosion with a structuring element of size 3 was used to reduce the overlap between the different classes. To cope with the fact that lacunae and canals had close grey levels, the MCW was applied by considering them as a single class. Then, the separation between these two was performed based on their size after a CCA. Each connected component was assigned a label, and the labels were sorted in decreasing order by component volume. Typically, canal formed a single large connected structure, but they could also consist of several disconnected structures due to volume cropping. Therefore, the canals were extracted by selecting the first labels (1–3). The number of labels for canals was selected manually based on visual feedback from 3D images for each sample. Finally, an initial segmentation of SCP, lacunae, canals and background was obtained using MCW and CCA. To automatically generated the total volume (TV), we utilized the logic operator OR to combine binary masks of the canals, SCP, and cell lacunae (as shown in Fig. 2).

3D nnu-net deep learning method

The nnU-Net architecture25 was employed for the automatic segmentation of cell lacunae (osteocytes and chondrocytes) in the SR phase contrast micro-CT images. Built on the standard U-Net architecture, the nnU-Net pipeline automatically managed preprocessing, training and post-processing tasks using Python and PyTorch. The computational blocks consisted of a series of convolutions, instance normalization, and leaky rectified linear unit activation functions. Downsampling (with increased number of feature channels) and upsampling (with decreased number of feature channels) were performed through strided and transposed convolutions, respectively. For incorporating additional features during the training, the architecture used skip connections or concatenations from the encoder side to the decoder side of the network. We set 250 training iterations for each epoch and trained the networks for 1000 epochs. Five-fold cross-validation was used to evaluate model performance, with cross-entropy loss and Dice loss as training cost functions.

To optimize the final segmentation of the cell lacunae, the logical operator AND was applied to eliminate over-segmentation caused by the MCW method and to reduce false positives caused by 3D nnU-Net. The final subchondral plate was then obtained using a subtraction operation, which removed canals and lacunae from TV (Fig. 2).

Generation of the training data

For training the 3D nnU-Net neural network, we need ground truth images with correctly labeled objects: SCP (SCB and CC), cell lacunae (osteocytes and chondrocytes), canals, and background. Typically, these labeled data are obtained through manual segmentation, which involves delineating the contours of objects and specifying the voxels of each compartment, in a slice-by-slice manner. Although manual segmentation is considered the ‘gold standard’ or ground truth for benchmarking automatic segmentation techniques, it requires expert knowledge and is often tedious and time consuming especially for 3D large datasets. In addition, manual segmentation is subject to intra and inter-observer variability26.

To improve the efficiency of generating annotated training data, we adopted a semi-automatic procedure involving manual correction of automatically segmented images. Initially, we applied the MCW method18 to segment SCP, canals, background and cell lacunae. However, over-segmentation of lacunae compartments was still present due to intensity variations and relatively low contrast around chondrocytes. We used the preliminary lacunae segmentation (Fig. S2 B, red box, Supplementary material) and corrected minor errors by hand to label the cell lacunae. Similarly, we manually checked and corrected misclassifications for the SCP, canals and background. To illustrate the process, the results obtained by using a manual method and the proposed semi-automatic method are shown in supplementary material Fig. S3.

In this study, we randomly selected and labelled 20 VOIs (\(\:2000\times\:1000\times\:100\)). Ten of these VOIs were used as training data for the 3D nnU-Net model, which was shown sufficient. The remaining 10 labeled VOIs were used to evaluate the segmentation quality. To ensure reliable evaluation and effective training, we selected sufficiently large slices \(\:(2000\times\:1000)\) to capture diverse image features, while using a relatively small number of slices (100) to minimize labeling time. The location of VOIs were manually determined based on visual feedback from the original images. The data information was listed in Table S1 in supplementary material.

Separation of chondrocytes and osteocytes

After segmenting the lacunae which included both chondrocyte and osteocyte lacunae, the main task was to separate these two types of cell lacunae.

Previous work on mice bone-tendon interfaces used differences in size and distance to the SCB to separate chondrocyte from osteocyte lacunae17. However, in our human knee samples, the sizes of chondrocytes and osteocytes were not always distinct. Additionally, the automatic extraction of SCB was challenging due to a frosted-glass appearance of CC compared to SCB with no variation of grey level intensity.

In this study, we used the CTAn micro-CT software (Bruker, Kontich, Belgium) to manually delineate the interface (cement line) between CC and SCB. This interface was visible due to the phase contrast in the images (Fig. S4, Supplementary material). Contours of the CC were generated every 20 slices, and interpolated to create the final CC mask, finally allowing the separation of chondrocytes from osteocytes. Generating these masks took approximately one hour per sample.

Evaluation of segmentation quality

Dice coefficient is a conventional statistical parameter to assess spatial overlapping between segmentation and the reference image27,28. Its expression is given by:

$$\:DSC=\frac{2\left|A\cap\:B\right|}{\left|A\right|+\left|B\right|}$$
(1)

where \(\:\left|A\right|\) and \(\:\left|B\right|\) are the cardinals of the two sets A and B, representing respectively the segmentation and ground truth.

To validate the segmentation quality, we calculated the Dice coefficient for each compartment (background, SCP, canals, and cell lacunae) by comparing the segmentation results obtained through a semi-automatic method to a reference image (watershed segmentation combined with manual correction). Ten independent VOIs from the training dataset and their labels (\(\:2000\times\:1000\times\:100\)) were used to evaluate the segmentation quality by calculating the Dice coefficients. To further evaluate segmentation quality and calculate sensitivity and specificity, we created the confusion matrix by calculating the spatial overlaps between the segmentation results and the reference image (generated by watershed + manual correction). For this evaluation, we did not separate chondrocyte and osteocyte lacunae, as the goal was to assess the automatic segmentation of cell lacunae; the separation of chondrocyte and osteocyte lacunae was performed manually using the CC mask. The data information is shown as Table S1 in supplementary material.

Quantitative parameters for analyzing subchondral plate, cell lacunae, and canals

Several parameters were extracted from the segmented SR phase contrast micro-CT images to quantitatively analyze SCP (CC and/or SCB), cell lacunae (chondrocyte and/or osteocyte lacunae), and canals. These parameters included calcified cartilage volume, volume fraction, lacunar number density, anisotropy and structure model index (SMI) of chondrocyte and osteocyte lacunae (Table 1). CC.TV is calcified cartilage total volume (Fig. S5 in supplementary material). The calculation of these parameters was based on a shell script and a homemade programs using the ITK library.

Table 1 Quantitative measured and calculated parameters.
Full size table

Volume-related parameters (CC.V, SCP.V/TV, Ca.V/TV, Lc.V/TV, Ch.V/CC.TV, and Oc.V/SCB.TV), in which total volume (TV) corresponds to the subchondral plate volume, were measured by counting voxels in each segmented compartment. Additionally, we calculated the lacunar number density for cell lacunae (Lc.N/TV), chondrocyte lacunae (Ch.N/CC.TV) and osteocyte lacunae (Oc.N/SCB.TV) using CCA method to count isolated particles. The 3D shape descriptors for individual chondrocyte lacunae: length (Ch.L1), width (Ch.L2) and depth (Ch.L3), and on osteocyte lacunae: length (Oc.L1), width (Oc.L2) and depth (Oc.L3) were calculated following a previously described method29. The ratios Ch.L1/Ch.L2, Ch.L1/Ch.L3, Ch.L2/Ch.L3 and Oc.L1/ Oc.L2, Oc.L1/ Oc.L3, Oc.L2/Oc.L3 were used to estimate the anisotropy of chondrocyte and osteocyte lacunae, respectively. The SMI was proposed to assess the rod-like or spherical nature of chondrocyte and osteocyte lacunae (Ch.SMI and Oc.SMI) with values of 3 for an ideal rod structure and 4 for an ideal spherical structure, and between 3 and 4 for a mixed structure of rods and spheres30. The SMI was defined as a function:

$$\:SMI=12\frac{V\cdot\:M}{{S}^{2}}$$
(2)

where \(\:V\) stands for volume, \(\:M\) represents integral of mean curvature, and \(\:S\) denotes surface area31.

Statistical analysis

The Lilliefors test and Bartlett’s test were used to test the normality and homoscedasticity of the quantitative parameters in each group (control or OA). All data were found to follow a normal distribution with equal variance. A t-test was used to determine whether there were significant differences in morphological parameters (calcified cartilage volume, volume fractions, and number densities of cell lacunae) between control and OA groups. A general linear mixed model was applied to compare parameters (volume, anisotropy, SMI of chondrocyte and osteocyte lacunae) between control and OA groups. Statistical analysis was conducted using the open source RStudio software 2022.12.0 with programming language R.

Results

Qualitative evaluation of segmentation

Qualitative examples of segmentation are illustrated in Fig. 3, showing 2D slices and zoomed-in regions of interest (A-E) and 3D renderings (F-H). The challenges for segmenting chondrocyte lacunae and canals are illustrated in Fig. 3 (A). Specifically, the CC surrounding chondrocyte lacunae results in relatively low contrast at the interfaces, and increases intensity variations between the different classes in the image. This prevents the MCW method from automatically selecting the correct markers, leading to the over-segmentation of chondrocyte lacunae (Fig. 3C, F), compared to the ground truth in Fig. 3 (B). Figure 3 (D, G) shows segmentation result using a nnU-Net method, which can learn the local features of various compartments and perform predictions with a trained model. Although the over-segmentation of chondrocytes was reduced, the thin and small canals were incorrectly segmented as lacunae (false positives) using a single nnU-Net method (Fig. 3D, G). This occured because the local features of the small canals were too similar to that of lacunae. Therefore, a single nnU-Net method may not be appropriate to provide a high-quality segmentation for canals. The combination of the MCW with the nnU-Net deep learning method addressed these challenges, as shown in Fig. 2. Qualitatively, the segmentation using the proposed protocol was significantly improved compared to the MCW method or the single nnU-Net method, as seen in Fig. 3 (E, H).

Fig. 3
figure 3

Representative segmentation examples. (A-E) 2D slices from 3D volumes. (F-H) 3D rendering of the segmented lacunae (in white) and canals (in yellow). (A) Original image. (B) Ground truth. (C)(F) MCW segmentation. (D)(G) nnU-Net segmentation. (E)(H) Segmentation using the combination of MCW algorithm and nnU-Net. The MCW method over-segmented lacunae. The single nnU-Net method incorrectly classified small and thin canals as lacunae. The proposed protocol provided well segmented canals and lacunae.

Full size image

Quantitative evaluation of segmentation quality

Ten representative VOIs (\(\:2000\times\:1000\times\:100\)) were used to validate the proposed protocol. First, the ten VOIs were segmented using the MCW method (M1), the nnU-Net method (M2), and the proposed protocol (M3), separately. Dice coefficients for the background, SCP, canals, and cell lacunae compartments were calculated on each VOI. Their mean and standard deviation values are shown in Fig. 4 (A). For SCP, canals, and cell lacunae, the specificity was 0.99 and sensitivity was ≥ 0.92. Details are shown in Table S2 (supplemental materials). According to the quantitative evaluation results, a relatively high segmentation quality was achieved using the proposed method. Significant improvements in lacunae segmentation were observed using the proposed protocol, compared to the MCW method alone. The segmentation of SCP, canals, and background was also noticeably improved using the proposed method compared to the single nnU-Net segmentation. Illustrative examples of segmentation are shown in Fig. 4 (B-F). The yellow arrows in Fig. 4 (D) highlight the over-segmented lacunae using MCW method, while the green arrows in (E) indicate the misclassification of canals and background compartments. However, a small portion of the SCP may have been misclassified as background during the watershed processing step.

Fig. 4
figure 4

Evaluation of segmentation quality. (A) Dice coefficients were calculated for various compartments using different segmentation methods. (B-F) Qualitative results obtained using different segmentation methods (all images are 2D slices from 3D volumes). The yellow arrow in (D) shows over-segmented chondrocyte lacunae using a MCW method. The green arrow in (E) highlights misclassifications of canals and background compartments. (F) demonstrates the superior performance of the combined MCW and nnU-Net method.

Full size image

Morphometric assessment of osteochondral interface

From the previously segmented images, quantitative parameters were calculated to compare the control group (ten samples) with the OA group (five samples). Descriptions of the parameters are listed in Table 1. Normal distributions and homoscedasticity of data were verified using the Lilliefors test (p-value > 0.05) and Bartlett’s test (p-value > 0.05), respectively.

The CC volume (CC.V), volume fractions (SCP.V/TV, Ca.V/TV, Lc.V/TV, Ch.V/CC.TV, Oc.V/SCB.TV), and lacunar number density (Ch.N/CC.TV, Oc.N/SCB.TV, Lc.N/TV) are listed in Table 2. Differences between group means and 95% confidence intervals are also provided in Table 2. No significant differences (at p-value < 0.05 level) were found between the control and OA groups. The anisotropy of chondrocyte lacunae (Ch.L1/Ch.L2, Ch.L1/Ch.L3, Ch.L2/Ch.L3) and Ch.SMI, as well as the anisotropy of osteocyte lacunae (Oc.L1/Oc.L2, Oc.L1/Oc.L3, Oc.L2/Oc.L3) and Oc.SMI are shown in Table 2. We found no significant differences between controls and OA for chondrocyte and osteocyte lacunae anisotropy, except for a decrease in the L2/L3 ratio (width to depth) in osteocytes lacunae from OA samples and a tendency towards a more spherical shape in the OA group. Using a general linear mixed model, the results for osteocytes anisotropy (L2/L3) and SMI were clearer with a statistically significant p-value = 0.048 for L2/L3 and nearly significant for SMI p-value = 0.058, as shown in Fig. S6 (Supplementary material).

Table 2 Statistical tests of parameters between control and OA groups (IQR: interquartile range, SD: standard deviation, CI: confidence interval, MD: mean difference). For comparisons of calcified cartilage volume, fraction, and number density, the t-tests were used, while for the parameters of cell lacunae (volume, anisotropy, SMI of chondrocyte and osteocyte lacunae), the general linear mixed models were applied.
Full size table

The primary orientations of chondrocyte and osteocyte lacunae in each bone core (with Y as the vertical axis and X-Z the horizontal plane) are shown in Fig. 5 and Fig. S7 (Supplementary material). We observed that the main orientation of most of chondrocyte lacunae in each sample was perpendicular to the cartilage surface in both control and OA groups. In contrast, the primary orientation of osteocyte lacunae appeared random, with similar proportions of orientations in both control and OA groups. This suggests that only chondrocytes exhibit elongated shapes, aligned vertically with the cartilage surface.

Fig. 5
figure 5

(A-B) The directions of X, Y, and Z axis in 2D slice and 3D original images. (C-D) The comparison of the main shape and orientation of chondrocytes between the control and OA groups.

Full size image

Discussion

In this study, we proposed a novel method to segment and quantify cell lacunae from 3D SR phase contrast micro-CT images of osteochondral unit of human knees. This method combines a MCW algorithm with a deep learning method based on nnU-Net.

Imaging cells in bone and cartilage is usually performed using light microscopy (LM) on stained histological sections. However, this approach provides only two-dimensional (2D) information limiting the assessment of the osteochondral interface and associated cells. Confocal scanning laser microscopy (CSLM), which produces three-dimensional (3D) reconstruction by stacking contiguous 2D images at varying depths within the specimen, is commonly used to image and characterize osteocytes with immunostaining32. However, CSLM is limited by the depth of light penetration and scattering when imaging mineralized tissues such as bone. Desktop micro-CT is a stain-free method offering volumetric virtual histology based on X-ray absorption contrast, has been used to study the 3D morphology of osteocyte lacunae in humans32,33. SR micro-CT possesses significant advantages over desktop micro-CT, in terms of image quality offering high spatial resolution, high signal-to-noise ratio and shorter data acquisition times. SR micro-CT has been used to image the osteochondral interface and quantify the 3D chondrocyte and osteocyte lacunae in both animal models15,17,34 and in humans29,35,36. Phase contrast imaging significantly enhances contrast and delineated borders compared to conventional methods and is able to discriminate different tissues with close absorption coefficients37. SR micro-CT coupled with phase contrast, has been implemented to study the osteocytes network16, cartilage35 or bone and cartilage at clinical resolution38 and to examine the bone-tendon interface15. This study represents the first application of synchrotron radiation CT with phase contrast at high resolution (0.65 μm) to quantitatively analyze cells morphology at the subchondral plate. Micro or nano-CT imaging is opening new possibilities for the precise characterization and quantification of tissue micro-architecture without labelling, potentially allowing larger sample sizes due to short acquisition time (5–10 min) and minimal sample preparation time39. Traditional two-dimensional cell counting methods and volume measurements based on histological approaches cannot serve as a reference for calcified tissues, as nano-CT with its isotropic voxels provides direct 3D measurements without the need of a stereological approach. We did not embed our samples, which were defatted by supercritical CO2, as cell sizes can be altered in paraffin-based histology due to the fixation process, which causes tissue shrinkage. However, histology remains indispensable for obtaining information on proteoglycan content and inflammation.

Segmentation is a crucial step in image analysis. Using simple threshold-based methods, we observed overlapping voxel intensity (grey level) between different structures (cell lacunae, canals, SCP), resulting in poor segmentation performance. Although the MCW method provided relatively convincing classifications, we observed over-segmentation of chondrocyte cell lacunae, likely due to the weak contrast at the interfaces between the lacunae and the incompletely calcified matrix. Deep learning methods have become increasingly popular for image segmentation. Here, we employed a nnU-Net deep learning method to segment cell lacunae, hypothesizing that a deep neural network could learn the local features of chondrocytes. Indeed, the trained model successfully distinguished chondrocyte lacunae from the surrounding incompletely mineralized matrix. However, the nnU-Net alone incorrectly classified small canal segments as lacunae, due to their similarity in their local features.

Our proposed protocol was compared qualitatively (Fig. 3) and quantitatively (Fig. 4) with the individual MCW and nnU-Net methods. Our results, based on Dice coefficient, demonstrated that the proposed protocol improved segmentation of each class. Specifically, it addressed the problem of separating partially calcified regions surrounding lacunae, which the MCW has failed to resolve. (Figures 3C and F and 4). The proposed method also reduced false positives in lacunae segmentation, which were previously caused by the similar local features of small canal in the single nnU-Net method (Figs. 3D and G and 4). SCP and background classifications were achieved using the MCW algorithm. Occasionally, the SCP was misclassified as background due to the three-dimensional flooding from background markers in different slices. The combined MCW and nnU-Net protocol successfully separated chondrocyte lacunae from the partially mineralized surrounding cavity in calcified cartilage of human knees.

The application of the proposed segmentation method to a real dataset of human knees yielded preliminary findings in both control and OA groups. In our study, the total volume of CC.V and SCB.V was 0.13 mm3 for CC and 0.34 mm3 for SCB, with on average 637 chondrocyte lacunae and 8474 osteocyte lacunae in the VOIs. Statistical analysis did not reveal significant differences in volume fractions (SCP.V/TV, Ca.V/TV, Lc.V/TV, Ch.V/CC.TV, Oc.V/SCB.TV) or cell lacunae density (Ch.N/CC.TV, Oc.N/SCB.TV, Lc.N/TV), likely due to the small sample size (Table 2).

According to the comparison in Oc.L2/Oc.L3 anisotropy in osteocytes between controls and OA, a tendency toward a more spherical shape in OA was found. Additionally, we found that the main orientation of most chondrocytes in each sample was vertical, a characteristic that was not clearly observed in osteocytes. While chondrocyte volume in CC has not been extensively studied, our preliminary results did not highlight differences between controls and OA. Chondrocyte hypertrophy in hyaline cartilage is known to play pivotal role in both early and late OA10. Changes in interstitial osmolarity and ionic composition of the pericellular matrix surrounding chondrocytes are known to influence chondrocyte volume, which increases under low osmolarity conditions typical of OA40. One possible explanation for the lack of volume change in CC chondrocytes is the presence of the surrounding extracellular matrix more or less calcified.

The pathophysiology of OA presents a challenge, as an extensive histopathological study suggested, that no discrete « normal » subgroup exists, indicating a pathological continuum between control and OA states41. In the present study, the control samples collected from elderly cadavers cannot be considered perfectly normal. In addition, we did not collect information about BMI, which is a potential bias, because no clinical data were available on anatomical subjects in accordance with regulatory authorities. Additionally, the classical radiographic Kellgren-Lawrence classification used in this study may be not appropriate for assessing the cartilage-bone interface as it tends to overemphasize the presence of osteophytes compared to joint space narrowing42. In two previous studies, we demonstrated that the central zone not covered by the meniscus already exhibited characteristics of high-load areas, which may explain the minimal differences observed between controls and OA43,44.

We have demonstrated the feasibility of studying the morphology and density of both chondrocytes and osteocytes. Without phase contrast, distinguishing and separately studying the CC and SCB, which contain two distinct populations of cells, would have been challenging.

Finally, the proposed method can be used to gain a better understanding of the roles of chondrocytes and osteocytes in SCP and may contribute to future studies of OA progression.