1. Introduction
Osteoarthritis (OA) is among the most common forms of arthritis [
1], occurring when protective hyaline cartilage between bones breaks down through injury or disease. Hyaline cartilage in the knee is an important tissue which, due to avascularity, does not heal spontaneously after injury and often requires surgical intervention. OA of the knee is a major cause of disability worldwide [
2], causing significant burden on healthcare systems [
3]. The lifetime risk of developing symptomatic knee OA is approximately 45%, with a higher risk being associated with obesity (60.5%) and advancing age [
4]. The incidence of OA is predicted to increase in the decades to come due to older populations and obesity [
5,
6].
OA is a polysymptomatic disease but is generally characterized by thinning and loss of cartilage. Assessment of cartilage condition and thickness is therefore crucial for both detection and monitoring the progression of OA. Diagnosis is usually based on a clinical assessment and a radiographic examination. Planar X-rays of the knee are routinely used in radiographic evaluation, however, soft tissue is not adequately visualized, nor is this modality sensitive to changes in the joint over time [
7]. Magnetic resonance imaging (MRI) is the state-of-the-art imaging modality for the assessment of hyaline cartilage and has seen rapid developments in hardware, sequences and image analysis in the last decade [
7,
8,
9,
10]. MRI provides a visual assessment of the cartilage and presents a means for quantitative evaluation of the volume and dimensions of the cartilage, and its chemical composition. Comprehensive overviews of MRI sequences for assessing morphological and compositional aspects of knee cartilage are described in [
7,
8,
11]. MRI is capable of accurately assessing the size and thickness of articular cartilage [
9,
12,
13]. In addition to visualizing the cartilage, MRI also images other tissues involved in OA, such as subchondral bone, meniscus and soft tissue. Computed tomography (CT) imaging also provides an excellent 3D representation of cortical bone [
14,
15,
16], osteophytes and soft tissue calcification and has been used to investigate changes in the joint, including trabecular bone remodeling, subchondral cysts and bone sclerosis, all of which can be OA-related changes in the joint [
17]. As the understanding of OA develops due to advances in medicine and imaging, it is important that OA is viewed as a disease of the whole organ, involving multiple joint tissues [
7].
The severity of OA can be assessed by the degree of joint space narrowing and damage to cartilage and underlying bone. Several scales that exist assess the extent of OA. Kellgren–Lawrence (KL) grading is used for the rating of OA on planar X-rays, where the definite presence of an osteophyte (Kellgren–Lawrence grade 2) confirms a structural diagnosis of OA [
18]. Kellgren–Lawrence combines an overall grade for OA from joint space narrowing and the presence of osteophytes, which incorrectly assumes that these structural changes appear continuously [
7]. Other grading systems such as the OA Research Society International (OARSI) Atlas system separate a joint space narrowing grade from the presence of osteophytes. Both, however, only assess the tibiofemoral joint, underestimating the patellofemoral contribution to the disease [
7]. Other commonly used scales include that developed by Ahlbäck in 1968 [
19] which is based on the measurement of joint space narrowing. A 2003 study measuring the inter- and intraobserver reliability of the Ahlbäck scale reported low to medium agreement coefficients, especially when reporting on radiographs of earlier stage OA [
20]. Comparisons of knee OA scales have shown moderate correlation with arthroscopic findings and also have moderate to high reliability between individual observers [
21]. In a study on severe OA, five radiological grading systems demonstrated medium correlation with intraoperative findings of full-thickness cartilage loss, and moderate interobserver reliability for all systems [
22]. In both studies, Kellgren–Lawrence and Ahlbäck showed the highest correlation with cartilage loss, although still in the moderate range. Semi-quantitative MRI-based grading systems such as the Whole Organ Magnetic Resonance Imaging Score (WORMS) and Knee OA Scoring System (KOSS) are based on a variety of features of the MR image from the whole knee joint, including cartilage size and depth, bone marrow lesions and subchondral cysts to name but a few. Some of these semi-quantitative scoring systems have demonstrated ‘within grade’ changes over time, thus exhibiting increased sensitivity compared to traditional grading systems [
7]. Quantitative measures using MRI include the cartilage volume and thickness calculated as a continuous variable, requiring segmentation of the cartilage in the MR image. This type of analysis requires high-resolution 3D imaging to facilitate accurate measurements and delineation of defined subregions of the cartilage [
7]. Comparison of several clinical, radiographic and biochemical measures revealed that the relatively strongest predictor of longitudinal MRI-defined cartilage thinning was reduced baseline cartilage thickness in the medial femur [
23]. Three-dimensional cartilage surface mapping (3D-CaSM) has also shown promise for tracking changes in cartilage thickness, where 6-month changes were observable using the semi-automatic 3D-CaSM algorithm [
24]. Recent work in healthy knees has also shown the utility of 3D ultrasound in quantifying cartilage volume when registered with MRI scans [
25].
Injury to the knee joint presents a substantial risk of development of OA. Planar X-ray imaging, CT arthrography and MRI have all been used in the assessment of knee trauma following injury [
26]. An MRI-based score incorporating traumatic and subsequent degenerative changes was introduced in 2014 by [
27]. The Anterior Cruciate Ligament OsteoArthritis Score (ACLOAS) evaluates structural joint damage, features of OA (including cartilage loss) and acute signs of inflammation in traumatic injury to the knee. The ACLOAS aims to be used for longitudinal assessment of injury and subsequent OA in the knee joint.
1.1. New Cartilage Assessment Methods/Gold Standard
There are two main volumetric analyses used in the cartilage assessment of this research. The first is a wall thickness analysis, where the cartilage mesh is taken and the thickness of each element is calculated from surface to surface. The hypothesis for this analysis is that patients’ degenerative and traumatic cartilages will be thinner in specific places based on the patient category than the control group. The second is the curvature analysis, which measures the Gaussian curvature of an element based on its surrounding elements. The hypothesis here is that around areas of higher cartilage degradation there will be higher curvature as holes and depressions form around these areas.
1.2. Use of 3D Modeling Tools
In this project, the medical 3D modeling software Materialise MIMICS (Materialise Interactive Medical Image Control System, Materialise, Belgium) and 3-Matic were used to analyze the cartilage. The application of the MIMICS software is to segment and isolate the specific cartilages from a CT scan and export the generated 3D mesh for further analysis. MIMICS allows the user to directly extract geometric measurements and densities in Hounsfield unit (HU) values for each element. The exported cartilage parts are subsequently imported into 3-Matic, where components’ meshes are analyzed and features extracted.
1.3. Machine Learning and Artificial Intelligence
In the scientific literature, it is possible to find multiple applications of machine learning (ML) and deep learning (DL) starting from MRI or other medical images of the knee related to OA. The progression of OA over time can be predicted with the use of ML algorithms using principal component analysis (PCA) [
28], X-ray images and general clinical data [
29]. DL was used by Liu et al. [
30] to detect acute cartilage injuries within the knee joint, while Bien et al. [
31] utilized DL efficiently to detect general abnormalities on knee MRI exams. Moreover, DL was used for OA diagnosis [
32] and the prevention of total knee replacement using MRI and non-image features [
33]. The KL grade system described above was predicted with DL by Kwon et al. [
34], having as initial features the gait data and radiographic images. Moreover, gene expression signatures and ML technologies were used to identify OA through liquid biopsy [
35,
36]. In this research ML, more specifically tree-based algorithms using novel 3D features proposed here, were implemented for a three-class classification.
This study presents a novel workflow developed in the frame of the EU project RESTORE. It is based on the segmentation and processing of medical images to 3D reconstruct a model of the knee joint. The analysis of these models provides an extensive set of metrics that can be used to assess cartilage and bone condition.
4. Discussion
This work aims to develop a new workflow to assess cartilage condition. Based on CT and MRI scans, a complete segmentation of the knee is performed. A 3D model is reconstructed from this segmentation, from which a unique set of data is extracted.
The relationship between systemic bone mineral density (BMD) and cartilage properties has been enhanced in recent years. Previous reviews indicate the positive correlation between BMD and cartilage defects, particularly related to the knee joint [
48,
49]. The results of this study show the degenerative group presenting lower BMD values compared to healthy and traumatic individuals. Although several studies have demonstrated that higher systemic BMD is associated with increased progression of cartilage defects, this relationship is still under investigation [
50]. A recent study demonstrated that BMD was the highest when the knee OA was mild, and was significantly lower in moderate and severe OA. Generally, it is considered that the severity of OA increases and the level of BMD decreases with increasing age [
51]. The different correlations between data could be due to the restricted number of participants, their characteristics, outcome measures or the status of the knee joint (as it is in the early stage). However, BMD could be an index of the cartilage condition because it is now validated that bone condition affects the course of the most common joint disease. In summary, the indexes based on the bone could show a way to differentiate the groups. This is proved by the conducted statistical analysis: for the patella bone, the D group can be discernible from C and T groups, and for the femur and tibia bones, it is differentiated from the T group. Moreover, the T group shows lower patella volume than the other groups. Some trauma may have involved the patella and caused its dislocation and subsequent cartilage lesion, leading to early OA [
52]. Thus, the examination of the patella volume can be of interest to investigate the presence of traumas.
Still, regarding the patella, its cartilage density is a good discriminator between C and D groups. In general, control patients present higher HU values for all cartilage segments, while those of traumatic and degenerative groups change from case to case. This could be because of the fact that cartilages in the early stages of OA generally present a greater amount of water with respect to physiologically normal cartilages [
53]. Therefore, the density calculated for the entire cartilage volume can discriminate between pathological and healthy conditions. Since traumatic and degenerative patients have opposite HU behaviors in lateral and medial tibial cartilages, while control ones remain constant for both parts, this result could represent proof of control regarding tissue density and pathologies.
The measurements of cartilages show a characteristic trend: volume increases for all the cartilages in degenerative patients. This is also strongly confirmed by the significant difference between the degenerative group and the other two groups for the femoral cartilage. Greater volume may indicate cartilage swelling in the early stage of degeneration due to an increased water content [
54,
55]. This is in agreement with biomechanical evidence, which suggests that the volume of degenerative cartilages and surrounding tissues would be greater than in healthy and traumatic circumstances. Moreover, all degenerative surface cartilage results were higher and we found a corresponding statistical difference between traumatic ones regarding the femoral cartilage. It is particularly interesting how the patella bone and cartilage play a fundamental role in the diagnosis. Until now, the evaluation of cartilage has been conducted mainly on the tibiofemoral joint, and the predominant scales for cartilage diseases do not take into account the patellofemoral compartment [
7]. Our study reveals that this patellar joint should be considered with more interest in knee assessment.
The visual inspection of the 3D models reveals that holes are present only in the degenerative cartilage. A third of the patients in the degenerative group present holes in their cartilage, mainly in the femur. Despite the limited amount of data analyzed in this paper, the fact that holes are only present in the degenerative group establishes groundwork for future research with aim of demonstrating that the detection of holes can be an indicator of degenerative cartilage.
The wall thickness analysis presented results of interest. A clear separation of the degenerative and traumatic patients from the control group is visible. Although a more significant separation was expected for the traumatic group, the wall thickness analysis did not take into account the holes that formed in the cartilages, which should be considered, as they are a form of degeneration. This is supported by Vincent and Wann [
56], as they have found a link between trauma to cartilage and a decreased wall thickness, as seen much more clearly in the degenerative case. The curvature analysis shows less clear results, although a separation for the degenerative group can be observed. Studies such as that of Folkesson et al. [
57] have found that the curvature for patients suffering from OA is significantly higher when compared to control patients, which is shown in the results here, as OA is a degenerative disease. Although it might be expected to find similar trends in the traumatic patients, the analysis results do not show as noteworthy a difference.
Most of the ML results can be considered highly satisfactory: it is clear that the models, especially using RF, can predict with good sensitivity the degenerative and traumatic subjects, and all the 3D novel features extracted appear to have a good predictive power (excluding the WT-C selection, whose results are lower). This can also be seen from the feature importance results (
Table 5), where cartilage and bone features assume a dominant role in the classification. To our knowledge, these features extracted manually from bones and cartilage images of the knee have never been used to distinguish the type of injury that provokes OA. A similar accuracy was obtained by Kwon et al. [
34] using gait and X-ray data (75.5%) predicting KL grade, while Du et al. [
28] using a PCA approach had a similar sensitivity with RF in the KL grade classification, achieving better results with other algorithms with respect to RF, such as simple artificial neural networks or support vector machines. A future improvement of the study can be to extend the classification to existing evaluation grades such as KL, which, knowing the current results, can give us interesting results in terms of accuracy and sensitivity and eventually other algorithms such as simple or advanced artificial neural networks can be used. A new index can also be developed, starting from these results, by integrating the 3D ones and other features extracted from the 2D image manual elaboration of the CT scans and MRI knee exams, for example, bone osteonecrosis, sclerosis, osteophytes and others.
Limitations
This work presents some limitations. The number of patients is limited. Moreover, the three categories are unbalanced: there are currently a lot of degenerative subjects and few control subjects. This affects the specificity results in the ML classification for the control subjects; in the future, the application of algorithms that can balance the number of subjects of the different classes (i.e., SMOTE [
58]) can be considered and, generally speaking, the results would be more accurate with an increased number of samples. More samples will eventually allow the use of more advanced artificial intelligence technologies such as DL and advanced neural networks. In addition, to improve the accuracy, the control patients could be excluded, and a binary classification can be performed on degenerative and traumatic patients.
The segmentation was carried out manually, leading to possible inaccuracies in the elements’ conformation. Although a standard protocol was defined, the overall procedure was performed by different people, and, occasionally, decisions were made based on visual interpretations. Therefore, some inaccuracies may be present in the final data.
Image quality also affects the initial steps of the segmentation process and the subsequent data extraction.
The wall thickness analysis also did not take into account any holes in the cartilage; this leaves out valuable information that could be extracted from the cartilage models as the traumatic patient group often had gaps in their cartilage. This could explain the relatively low percentage of elements below the standard deviation of the mean in
Figure 7 compared to the degenerative patients.