Characterisation of disease heterogeneity in malignant pleural mesothelioma using mixture modelling of ADC and R2
Lin Cheng1, Matthew D. Blackledge1, David J. Collins1, Nina Tunariu1,2, Neil P. Jerome1, Matthew R. Orton1, Veronica A. Morgan3, Martion O. Leach1, and Dow-Mu Koh1,2

1Division of Radiotherapy and Imaging, Cancer Research UK Cancer Imaging Centre, Institute of Cancer Research, London, United Kingdom, 2Radiology, Royal Marsden Hospital, London, United Kingdom, 3Clinical MRI Unit, Royal Marsden Hospital, London, United Kingdom

Synopsis

Disease heterogeneity in patients with malignant pleural mesothelioma (MPM) makes it challenging to characterise solid disease and assess response following treatment. Computed Diffusion-Weighted MRI (cDWI) provides improved contrast between disease and background tissues, and facilitates total disease segmentation. A mixture modelling of ADC and R2 with semi-automatic segmentation on the cDWI is proposed to assess disease heterogeneity in MPM, with demonstration of its utility on a paired pre/post-treatment dataset. The mixture modelling methodology successfully characterised disease heterogeneity for two MPM patients, and can provide additional quantitative functional disease response characterisation compared with using only a single parameter.

Introduction

Disease heterogeneity in patients with malignant pleural mesothelioma (MPM), which presents with both solid tumour and pleural effusion, makes it challenging to characterise solid disease and assess response following treatment. Computed Diffusion-Weighted MRI (cDWI) provides improved contrast between disease and background tissues, and facilitates total disease segmentation [1]. Mixture modelling of joint histograms of within-tumour Apparent Diffusion Coefficient (ADC) and transverse relaxation rate (R2) may offer a viable method to charactierise disease heterogeneity. The purposes of this study were to investigate this novel methodology to assess disease heterogeneity in mesothelioma, and demonstrate its utility on a paired pre/post-treatment dataset.

Methods and Materials

MR image acquisition: Diffusion- and T2-weighted images were acquired in five patients with MPM, before and 4 weeks after anti-tumour treatment, using 1.5T Magnetom Avanto (Siemens Healthcare, Erlangen, Germany). A 2-station EPI sequence (covering the whole chest) was used for both acquisitions. The DWI parameters were: 30 axial slices, slice thickness 5mm, TR/TE=9000/82 ms, FOV=273*380 mm, 4 averages, matrix 92*128, parallel acquisition (Grappa acc. factor 2, ref lines 30), receiver bandwidth = 1860Hz/Px, fat suppression SPAIR, b =100/500/800 mm-2s. T2-weighted images were acquired using the same sequence and matrix, with b-value of 0 and TEs 60/82/177 ms.

Segmentation/ mixture modelling: The total disease volume was segmented using a semi-automatic tool (in-house software based on the GrowCut algorithm [2]) using cDWI images with b = 150 s/mm2 and echo time 20ms (in-house software [1]), chosen such that the two components of MPM had similar signal intensities, whilst maintaining good disease/background contrast. Segmented ROIs were reviewed by a senior radiologist and then transferred to calculated ADC and R2 maps, jointly estimated from the diffusion- and T2-weighted data [1]. Two-dimensional scatter plots of ADC/R2 voxel values were generated and a 2-class Gaussian Mixture Model (GMM) was applied to the global multivariate data, using the Expectation Maximization (EM) algorithm for parameter estimation [3]. Following estimation of these multi-modal probability distributions, deriving the posterior probability, Pij, of a voxel i belonging to each class Cj, allowed construction of tissue classification maps and calculation of proportional disease volume of each class.

Results

Patient 1: It can be seen (Figure 1) that the two pre-treatment baseline data have very similar shapes of ADC/R2 distribution on the joint histogram and are both well fitted by GMM (contour lines) with two Gaussian distributions (mean vectors shown on the figure). The 3rd column of Figure 1 displays the regions of solid disease (red) and pleural effusion (green) on the ADC maps, while the 4th column shows maps of the posterior probabilities for the two classes (C1: solid disease; C2: pleural effusion). In addition, regions and their probability maps show a high correspondence between the two baseline studies. The proportional volumes of the two classes have < 1% difference in the two pre-treatment baseline scans (Patient 1 in Table 1), suggesting a repeatable method, despite an apparent difference in centroid ADC for C1.

Patient 2: After 4 weeks treatment, the joint parameter distribution changed with increased dispersion, with increasing scatter observed on the left corner of the joint histogram, representing post-treatment changes in class 1 (lower ADC and lower R2) (Figure 2). There was a visible shift of the centre of cluster 1 (solid disease) to the lower left, ADC from 1538 to 1411 s/mm-2 while R2 from to 13.6 to 12.3 s-1 (Figure 2 and Table 1). After treatment, both the absolute and proportional volume of class 1 (solid disease) increased while the pleural effusion decreased (Patient 2 in Table 1).

Discussion

In these two cases, we have demonstrated that it is possible to assess disease heterogeneity by using joint ADC/R2 modelling combined with a semi-automatic segmentation tool on cDWI. This method can be used to observe changes in the mean ADC and R2 values after treatment within a disease volume for each tissue class independently and evaluate the heterogeneous treatment response. Further investigation of the model will be performed in a larger clinical cohort.

Conclusion

The mixture (ADC and R2) modelling methodology successfully characterised response heterogeneity for two patients with malignant pleural mesothelioma and can provide additional quantitative functional disease response characterisation compared with using only a single parameter.

Acknowledgements

CRUK and EPSRC Cancer Imaging Centre in association with the MRC and Department of Health grant C1060/A10334; NHS funding to the NIHR Biomedical Research Centre and post-doctoral fellowship funding by the NIHR (NHR011X); An Experimental Cancer Medicine Centre Network award (C51/A7401 & C12540/A15573); British Lung Foundation grant (APP13-4). Martin O Leach is a senior NIHR investigator.

References

[1]. Cheng et al., Proc 23rd Annual Meeting ISMRM 2015, 3599

[2]. Vezhnevets, V. and Konouchine V., Proc. Graphicon, 150-156, 2005

[3]. Pedregosa et al., JMLR 12, 2825-2830, 2011.

Figures

Figure 1: A 68-year-old male MPM patient with two pre-treatment scans (one week apart). Columns (left to right): R2 maps; ADC maps; ADC maps with overlaid classes: solid tumour (C1, Red); pleural effusion (C2, Green); posterior class probability map; scatter plots showing properties of the two clusters (colours represent maximum a-posteriori classification for each voxel).

Table 1: ADC, R2 and T2 values for the cluster centres in Gaussian distributions and the absolute and percentage volumes for two patients.

Figure 2: A 61-year-old female MPM patient with pre- and post-treatment scans 4 weeks apart. Columns (left to right): R2 maps; ADC maps; ADC maps with overlaid classes: solid tumour (C1, Red); pleural effusion (C2, Green); posterior class probability map; scatter plots showing properties of the two clusters (colours represent maximum a-posteriori classification for each voxel).



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
2432