Dominik Vilimek1, Veronika Janacova2,3, Radana Vilimkova Kahankova1, Pavla Hanzlikova4,5, Jindrich Brablik1, Michaela Pomaki4,5, Siegfried Trattnig2,3,6,7, Radek Martinek1, and Vladimir Juras2
1Department of Cybernetics and Biomedical Engineering, VSB Technical University of Ostrava, Ostrava, Czech Republic, 2Department of Biomedical Imaging and Image-guided Therapy, High Field MR Centre, Medical University of Vienna, Vienna, Austria, 3CD Laboratory fo MR Imaging Biomarkers (BIOMAK), Vienna, Austria, 4Department of Imaging Methods, Faculty of Medicine, University of Ostrava, Ostrava, Czech Republic, 5Department of Imaging Method, University Hospital Ostrava, Ostrava, Czech Republic, 6Austrian Cluster for Tissue Regeneration, Ludwig Boltzmann Institute for Experimental and Clinical Traumatology, Vienna, Austria, 7Institute for Clinical Molecular MRI in the Musculoskeletal System, Karl Landsteiner Society, Vienna, Austria
Synopsis
Keywords: Cartilage, Radiomics
Motivation: Our study examines the utility of graph-based analyses in revealing the interplay of radiomic features in knee osteoarthritis (OA), specifically to discover patterns that are hidden in traditional analyses.
Goal(s): To differentiate radiomic profiles of healthy individuals from OA patients using graph-based methodologies and identify key features associated with OA progression.
Approach: We analyze feature interconnections within knee joint compartments using MRI-based radiomics and cosine similarity graphs to evaluate features from 20 subjects.
Results: Clustering coefficients and path lengths within the graphs revealed a distinct, pathology-driven convergence of radiomic features in OA patients compared to controls.
Impact: The graph analysis
revealed a convergence of radiomic features in OA, potentially contributing to
a better understanding of the disease and therefore opening the path to novel
analysis strategies.
Introduction
The rapidly developing field of
radiomics has opened up another frontier in medical imaging. As a result, itis
possible to extract a wide range of quantitative features from standard
clinical images in a high-dimensional manner. These features, which capture the
underlying pathophysiology of tissues, have shown promise for improving
diagnosis, prognosis, and treatment stratification in a variety of diseases,
including OA1,2,3. Radiomics offers tens of
features, such as shape, texture, and intensity, that provide a comprehensive
understanding of the image4,5. However, other methodologies are
needed to understand not just the individual features but also their
interactions. Thus, in this study, we use graph-based methods to visualize and
analyze complex interconnections between MRI-based radiomic features, providing
a system-level perspective that is often obscured in traditional analysis. This
graph-based methodology not only enhances our understanding of the radiomic
landscape in knee OA but also sets the stage for developing predictive models
that could transform patient management by providing
a nuanced, data-driven view of the disease.Methods
Analysed cohort consisted of 20 subjects from the baseline
measurement of the Osteoarthritis Initiative (OAI) study. Selected subjects
did not have knee injury or surgery prior to the baseline measurements. Ten
subjects (5 males, 5 females, 49.8±1.7 years, BMI 32.0±1.4, KL (Kellgren and
Lawrence system) =0) served as controls. Other ten subjects (4 males, 6
females, 51.2±2.6 years, BMI 32.3±1.1, 5x KL=2, 5x KL=3) were considered as
patients with OA.
The automatic segmentation was performed on the Double Echo
Steady State images (3D DESS) using the MR ChondralHealth version 3.1
research application software (Siemens Healthineers AG, Forchheim, Germany,
segmentation algorithm based on the work from6,7) obtaining 9
anatomical regions for the femoral cartilage: medial anterior (MaF)/central
(McF)/posterior (MpF); trochlear lateral (TL)/central (TC)/medial (TM) and
lateral anterior (LaF)/central (LcF)/posterior (LpF), see Figure 1.
Utilizing the PyRadiomics
library4, we extracted a suite of twenty features including shape, size,
texture, and intensity-based parameters for each segmented label. Feature
extraction was parameterized to include a comprehensive set of Gray Level
Co-occurrence Matrix (GLCM) features, capturing the textural patterns within
the cartilage that may correlate with early degenerative changes (e.g. autocorrelation, contrast or entropy).
Then, we grouped the radiomic features based on
their corresponding anatomical regions within the knee joint: medial, lateral,
and patellofemoral compartments. We normalized the features to account for
inter-patient variability and employed cosine similarity to construct graphs,
with nodes representing individual patients and edges weighted by feature
vector similarities. Community detection algorithms and centrality measures
were employed to discern the structure, aiming to identify
patterns indicative of healthy versus pathological states. The Girvan-Newman
algorithm was used for community detection, while degree, closeness, and
betweenness centrality metrics provided insights into the connectivity and
importance of individual nodes.Results
The similarity graphs
revealed distinct clustering patterns between control and OA group. In the
control, the graph had sparser connections, reflecting a more diverse
radiomic feature set. Conversely, the patient group graph was characterized by
denser connectivity, suggesting a mutual convergence of features due
to pathological changes, see Figure 2.
Interestingly, the Girvan-Newman method identified three communities within the
graph.
There were pronounced differences in the lateral and medial
compartments between the groups, with the patient graphs showing
greater clustering coefficients and shorter average path lengths, see Figure 3.
The graph properties
reveal the average clustering = 0.523 and 0.713 for control and
patient group, respectively. There was a slight increase in the average
shortest path length in the patient group (1.956) compared with the control
group (1.667). A higher global clustering coefficient is found in the patient
group (0.692) than in the control group (0.484). The centrality analysis
identified several nodes with a high betweenness centrality in the patient
group; these may represent pivotal patterns in the radiomics associated with OA progression.Discussion and Conclusion
In our study, a graph-based analysis of radiomic features
in knee OA revealed a nuanced picture of the disease's impact on the
knee's properties. The observation of denser connectivity in the patient
group compared to controls suggests that OA induces a mutual convergence of
radiomic features, reflecting underlying pathophysiological changes.
Therefore, graph-based methods can enhance our understanding of OA. The study has several limitations, such as the limited
number of patients and the limited number of features. Further research should investigate the use of graph-based methods for other prominent quantitative MRI
techniques such as T2 mapping or dGEMRIC. Besides investigating
the therapeutic implications of the detected patterns, these
potential biomarkers require further validation. Thus, as radiomics advances, graph-based analyses will be more important for interpreting high-dimensional imaging data.Acknowledgements
This article was co-funded by the European Union under
the REFRESH – Research Excellence For Region Sustainability and High-tech
Industries project number CZ.10.03.01/00/22 003/0000048 via the Operational
Programme Just Transition and by the Ministry of Education of the Czech
Republic under project SP2023/042.References
- Janacova, V., Szomolanyi,
P., Kirner, A., Trattnig, S., & Juras, V. (2022). Adjacent
cartilage tissue structure after successful transplantation: a quantitative MRI
study using T2 mapping and texture analysis. European Radiology, 32(12),
8364-8375.
- Hirvasniemi, J., Klein, S.,
Bierma-Zeinstra, S., Vernooij, M. W., Schiphof, D., & Oei, E. H. (2021). A
machine learning approach to distinguish between knees without and with
osteoarthritis using MRI-based radiomic features from tibial bone. European
radiology, 31, 8513-8521.
- Lin, T., Peng, S., Lu, S.,
Fu, S., Zeng, D., Li, J., ... & Ding, C. (2023).
Prediction of knee pain improvement over two years for knee osteoarthritis
using a dynamic nomogram based on MRI-derived radiomics: a proof-of-concept
study. Osteoarthritis and Cartilage, 31(2), 267-278.
- Van Griethuysen, J. J.,
Fedorov, A., Parmar, C., Hosny, A., Aucoin, N., Narayan, V., ... & Aerts,
H. J. (2017). Computational radiomics system to decode the radiographic
phenotype. Cancer research, 77(21), e104-e107.
- Rizzo, S., Botta, F.,
Raimondi, S., Origgi, D., Fanciullo, C., Morganti, A. G., & Bellomi, M.
(2018). Radiomics: the facts and the challenges of image analysis. European
radiology experimental, 2(1), 1-8.
- Fripp, J., Crozier, S.,
Warfield, S. K., & Ourselin, S. (2009). Automatic segmentation and
quantitative analysis of the articular cartilages from magnetic resonance
images of the knee. IEEE transactions on medical imaging, 29(1), 55-64.
- Chandra,
S. S., Xia, Y., Engstrom, C., Crozier, S., Schwarz, R., & Fripp, J. (2014).
Focused shape models for hip joint segmentation in
3D magnetic resonance images. Medical image analysis, 18(3), 567-578.