MR texture analysis of subchondral bone in osteoarthritis
James MacKay1, Samantha Low1, Philip Murray1, Bahman Kasmai1, Glyn Johnson2, Simon Donell2,3, and Andoni Toms1,2

1Radiology, Norfolk & Norwich University Hospital, Norwich, United Kingdom, 2Norwich Medical School, University of East Anglia, Norwich, United Kingdom, 3Trauma & Orthopaedics, Norfolk & Norwich University Hospital, Norwich, United Kingdom

Synopsis

Subchondral bone (SB) plays an important role in osteoarthritis (OA). Texture analysis (TA) is a method of quantifying changes in SB and may be a useful OA biomarker. The optimum MR method to allow TA of SB is unclear. We compared TA using two sequences which demonstrated promise in depicting SB structure in three groups of participants: normal controls (n=10), individuals with early OA (n=10) and individuals with advanced OA (n=10). TA using a 2D T1-weighted spin echo sequence demonstrated more significant differences in texture features between groups and improved classification accuracy compared to a 3D gradient echo sequence.

Purpose

To compare the ability of MR texture analysis (TA) using 2D T1 weighted spin echo (T1SE) and 3D gradient echo (GRE) images to quantify subchondral bone (SB) changes in knee osteoarthritis (OA).

Methods

30 participants were recruited in three groups of 10. The first group (A) contained asymptomatic volunteers aged between 20 – 30 years old with normal body mass index. The second group (B) contained individuals aged between 40 – 50 years old who had been referred to the Orthopaedic department at our institution for evaluation of knee pain, with no radiographic evidence of OA (Kellgren-Lawrence grade ≤ 1). The third group (C) contained individuals scheduled to undergo total knee replacement.

Participants underwent MR imaging of the knee on a wide-bore 3T MR platform (GE 750w) using an 8-channel knee coil (GE WD 750). To evaluate the SB, we performed a 2D coronal T1SE sequence (FOV 12 x 12.3 cm, matrix 512 x 512, TR 593 mSec, TE 17.65 mSec, NEX 1, slice thickness 2.5 mm, interslice gap 2.8 mm, scan duration 2 mins 35 secs) and a 3D coronal GRE sequence (FOV 12 x 12.3 cm, matrix 512 x 512, TR 11.58 mSec, flip angle 50o, TE 4.33 mSec, NEX 0.6, slice thickness 1 mm, scan duration 3 mins 45 secs) (figure 1).

Six T1SE coronal images and the corresponding six GRE images through the central weight-bearing portion of the tibial plateau were selected. Regions of interest (ROI) were created in the medial (MT) and lateral (LT) tibial SB (figure 2). Identical ROIs were used for T1SE and GRE images. For each ROI, 20 statistical texture features (table 1) were calculated using dedicated TA software (MazDa v4.6)1.

The mean value of each calculated texture feature was compared between groups using a one-way ANOVA with post-hoc t-tests where a significant difference between groups was demonstrated. The Bonferroni method was used to adjust for multiplicity of testing, with p < 0.0025 (0.05/20) considered statistically significant. The number of features different between groups using T1SE and GRE images was compared.

We created two exploratory linear discriminant functions using the most discriminating T1SE and GRE texture features in order to assess the ability to classify participants into the correct group,. The classification accuracy of each function was then assessed using linear discriminant analysis (LDA).

Results

TA using 2D T1SE images demonstrated significant differences between groups in 18/20 texture features at the MT and 12/20 texture features at the LT. At the MT, 2 features were significantly different between groups A & B, 13 features were significantly different between groups B & C, and 17 features were significantly different between groups A & C. At the LT, 3 features were significantly different between groups A & B, 7 features were significantly different between groups B & C, and 12 features were significantly different between groups A & C.

TA using 3D GRE images demonstrated significant differences between groups in 17/20 texture features at the MT and 11/20 features at the LT. At the MT, no features were significantly different between groups A & B, 14 features were significantly different between both groups B & C and groups A & C. At the LT, no features were significantly different between groups A & B, 3 features were significantly different between groups B & C, and 11 features were significantly different between groups C & D (table 2).

LDA using texture features extracted from T1SE images classified 29/30 subjects into the correct group (97% accuracy, 95% CI 91-100) for both MT and LT datasets. LDA using texture features extracted from GRE images classified 21/30 subjects into the correct group (70%, 54-86) using MT data and 19/30 subjects into the correct group (63%, 46-81) using LT data (figure 3).

Discussion

TA using both 2D T1SE and 3D GRE images demonstrated significant differences in SB texture between the three groups. More texture features were significantly different between groups using 2D T1SE images, and classification of subjects was better at both MT and LT using 2D T1SE data. The apparent superiority may be due to the greater slice thickness of the 2D T1SE sequence providing higher SNR, and less susceptibility artefact arising from the subchondral trabeculae when compared to the 3D GRE sequence.

Conclusion

MRI TA of SB using 2D T1SE is superior to TA using 3D GRE. TA offers a promising method of assessing alterations in SB at different stages of OA.

Acknowledgements

This study was funded by a Royal College of Radiologists' Pump Priming Grant. The research team acknowledge the support of the National Institute for Health Research, through the Comprehensive Clinical Research Network. We would like to thank Angela Bullough and Sue Butters, Orthopaedic Research Nurses, for their assistance with participant screening and recruitment.

References

1. Szczypinski P, Strzelecki M, Materka A, Klepaczko A. MazDa - a software package for image texture analysis. Comput Methods Programs Biomed. 2009;94(1):66-76.

Figures

Figure 1: Sample (A) 3D gradient echo and (B) 2D T1 spin echo coronal images of the knee of a group A subject

Figure 2: ROI creation in the medial tibial subchondral bone (white dashed line).

Table 1: Summary of texture features calculated

Table 2: Number of texture features significantly different between groups using 2D T1SE and 3D GRE images (total number of features compared = 20)

Figure 3: Results of linear discriminant analysis at the medial and lateral tibia using 2D T1SE and 3D GRE images. Subject classification was superior with 2D T1SE images as demonstrated by improved separability of the data points. Each data point is a single ROI (n=180).



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
4485