Texture analysis of T2 relaxation time maps reveals degenerative changes in articular cartilage: Oulu Knee Osteoarthritis study
Arttu Peuna1,2,3, Joonas Hekkala1,3, Marianne Haapea1,2, Jana Podlipska1,3, Ali Guermazi4, Miika T Nieminen1,2,3, Simo Saarakkala1,2,3, and Eveliina Lammentausta1,2

1Medical Research Center, University of Oulu and Oulu University Hospital, Oulu, Finland, 2Department of Diagnostic Radiology, Oulu University Hospital, Oulu, Finland, 3Research group of Medical Imaging, Physics and Technology, University of Oulu, Oulu, Finland, 4Department of Radiology, Boston University School of Medicine, Boston, MA, United States

Synopsis

Gray level co-occurrence matrix based texture analysis is a sensitive image processing method that probes the spatial information from knee MR T2 maps and of the changes caused by osteoarthritis (OA). Texture analysis can distinguish symptomatic patients from healthy control subjects more sensitively than regional mean T2 analysis, and provides additional information also when compared to clinical evaluations such as MOAKS. Advanced learning algorithms can be further utilized to classify asymptomatic and OA subjects.

Introduction:

Texture analysis methods based on gray level co-occurrence matrices (GLCM)1, are image processing tools that can extract spatial correspondence information from digital images otherwise uninterpretable by human perception. T2 relaxation time mapping has proven promising in assessing cartilage quality2. GLCM analysis can be used to probe underlying information from T2 maps and of changes caused by osteoarthritis (OA), thus leading to a more sensitive characterization of osteoarthritic cartilage.

In this abstract, we present how texture analysis is able to distinguish OA patients from apparently healthy controls using various texture parameters and show how advanced learning algorithms, K-nearest neighbors (KNN) and multilayer perceptron learner (MPL), perform on texture data.

Materials and Methods:

64 asymptomatic volunteers (38 female, mean age 54.9 years (standard deviation, SD 13.9), body mass index 24.9 kg/m2 (SD 3.2)) and 80 symptomatic patients (49 female, 59.9 years (SD 7.7), 29.0 kg/m2 (SD 4.29)) were recruited with the permission from the local ethics committee. In total 80 asymptomatic volunteers were recruited via newspaper advertisement, while 16 subjects had to be excluded due to clinical MRI OA findings. For the patient group, subjects were selected from the hospital registry based on a long term non-specific knee pain or referring to knee replacement surgery. Exclusion criterion included trauma, rheumatoid diseases, previous knee surgery, or other underlying medical conditions that could affect the knee joint.

Subjects were scanned on 3T Siemens Skyra (Siemens Healthcare, Erlangen, Germany). T2 relaxation time mapping was performed with a multi-slice multi-echo spin echo sequence (TR=1680ms, TE’s(5)=13.8…69.0ms, ETL=5, FOV=160x160mm, matrix=384x384, thickness=3 mm). ROI-wise texture analysis was applied for segmented T2 maps with an in-house developed segmentation and analysis tool (MATLAB, Mathworks Inc., Natick, MA). MOAKS scoring3 was performed from clinical MRI sequences by an experienced radiologist.

For evaluating MPL, 45% of subjects were used for training, 25% for validating and 30% for testing of the MPL. KNN was trained with 65%/35% training/testing partitions with Euclidean distance cost. For both learning tasks, 264 texture variables were used (3 offset directions 22 parameters in each for 4 different ROIs). Statistical analysis was performed using IBM SPSS statistics (IBM Corp., Armonk, NY).

Results:

Texture parameters displayed excellent performance in discerning patients with clinical OA and asymptomatic volunteers (Table 1). Whereas mean T2 values showed significant difference between control and patient group only in lateral central femoral region, difference entropy, contrast, homogeneity and information measure of correlation are significantly different in all studied regions and energy in femoral regions. Texture parameters also provided smaller p-values for lateral central femoral region as compared to mean T2.

The same texture variables were studied in subsets of patients and controls without cartilage related findings in ROI’s MOAKS (cartilage loss, cartilage full thickness loss or osteophytes). In other words, only the subjects with MOAKS score of zero in corresponding ROI were included. These groups show similar outcomes to the whole dataset apart from medial central femoral region, where significant difference is displayed only on difference entropy, contrast and homogeneity (Table 2).

MPL showed moderate to good performance in classifying dataset into patients and controls (Table 3). Correct classification rate for control subjects was 14 out of 19 (73.7%) and for patients 15 out of 17 (88.2%). ROC curves (Figure 1) shows excellent accuracy for the classification results (Area Under the Curve = 0.950 for both the controls and patients). KNN correctly classified 22 out of 25 (88.0%) controls and 19 out of 25 (76.0%) patients into right class (Table 4).

Conclusion:

Texture analysis appears more sensitive to cartilage degeneration than mean T2 when the same relaxation time maps are used for both analyses. Relaxation time maps contain underlying information that cannot be interpreted with simple statistics or visual interpretation, however, with GLCM it is possible. When compared to clinical evaluations, texture analysis shows high performance: even when MOAKS scores indicate no radiological changes in the cartilage, analyzing T2 maps with texture analysis method allows discerning patients from healthy volunteers. MPL and KNN both have good performance and allow advanced analysis of big texture datasets.

Acknowledgements

No acknowledgement found.

References

1. R. M. Haralick, et al., "Textural Features for Image Classification," IEEE Transactions on Systems, Man, and Cybernetics, vol. 3, no. 6, pp. 610-621, 1973.

2. F. Eckstein, et al., "Quantitative MRI of cartilage and bone: degenerative changes in osteoarthritis," NMR in Biomedicine, vol. 19, pp. 822-854, 2006.

3. D. J. Hunter, et al., "Evolution of semi-quantitative whole joint assessment of knee OA: MOAKS (MRI Osteoarhritis Knee Score)," Osteoarthritis and Cartilage, vol. 19, pp. 990-1002, 2011.

Figures

Table 1. Mean T2 and texture analysis values for different regions of interests. Number of controls was 64 and number of patients was 80.

Table 2. Mean T2 and texture analysis values in a subset of patients and controls without ROI-correlated cartilage related findings in MOAKS. Numbers of controls/patients were 41/15 for medial central femur, 57/36 for lateral central femur, 51/26 for medial central tibia and 50/33 for lateral central tibia.

Table 3. Multilayer perceptron learner shows good overall accuracy when classifying OA patients and healthy controls. 264 texture parameters were used for learning. 45% of subjects were used for training, 25% for validating and 30% for testing of the algorithm.

Figure 1. ROC performance of multilayer perceptron learner. Area Under the Curve is 0.950 for both the control and patient curves.

Table 4. K nearest neighbor classifier shows good overall classification accuracy. 264 texture parameters were used for learning and 65% of subjects were used in training and the remaining 35% in testing of the algorithm.



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
0538