Longitudinal Changes of White Matter Lesions

INTRODUCTION: Quantification of white matter lesion progression is of major interest in MR imaging of the human brain for the diagnosis and monitoring of several diseases like Multiple Sclerosis or Alzheimer’s disease. T1 (T1w) and T2 (T2w) weighted images may provide insufficient tissue contrast for accurate segmentation of lesions. The Fluid Attenuated Inversion Recovery (FLAIR) sequence is often used to address this problem, as it provides superior contrast for the detection of lesions. Unfortunately, FLAIR images are not always available, especially in retrospective studies involving older MR data sets. In this work, we synthesize FLAIR contrast images from corresponding T1w and T2w images, enabling lesion detection consistent with that achieved with T1w + FLAIR images. Using our synthetic FLAIR images, we quantify progression of lesions in subjects from the Baltimore Longitudinal Study on Aging [3], which only possess FLAIR sequences during later time points. METHODS: We build upon the idea of image hallucination [1] to synthesize FLAIR images from a set of T1w and T2w images. Suppose fT1 and fT2 are T1w and T2w images of a subject. If fFL is the FLAIR modality, then, fFL = W(fT1, fT2,θ), where W is the imaging process, θ is the intrinsic acquisition parameters like repetition time, flip angle etc. Given all these underlying parameters, it is possible to directly estimate fFL [4]. For most studies, θ is not known and W is difficult to model, prohibiting direct estimation of a FLAIR. Instead we use an atlas based approach to synthesize FLAIR. We assume we are provided or can obtain an atlas, defined as a set of co-registered images A={gT1,gT2,gFL,hL}, acquired from a model subject that is different from f but also possesses lesions, where gT1, gT2, gFL and hL are T1, T2, FLAIR images and hard segmentations of lesions, respectively. Also assume that the images are composed of 3D patches fT1(i), fT2(i), gT1(j), gT2(j), gFL(j), centered at Ω , Ω voxels where Ωf and Ωg are image domains of f and g respectively. We synthesize the FLAIR image FL by matching patches of fT1 and fT2 to the atlas gT1 and gT2 , taking the resulting patch from gFL. Fig. 1 visually depicts the algorithm flow. From the subject f, we take the th i patch pair { fT1(i), fT2(i)} and identify the best possible matching pair {gT1(j), gT2(j); j Ω }.The corresponding FLAIR patches {gFL(j), j Ω } are combined using a non-local means approach [5] to generate the synthetic FLAIR patch FL . The merging of all such patches generates the synthetic FLAIR FL . RESULTS: We used this method to synthesize FLAIR images for a longitudinal dataset, which has only T1w and T2w images for older time points and FLAIR for newer time points. We use a validated multichannel lesion segmentation tool [2] to find lesions. To validate the synthetic FLAIR image, we use the lesion segmentation from T1w and true FLAIR as a reference standard and compare the segmentation from T1w and synthetic FLAIR with this reference standard. The Dice between the segmentation obtained from T1w+T2w (Fig. 2) with the reference is 0.51, while it is 0.69 between the reference and the segmentation obtained from T1w+synthesized FLAIR (Fig. 2). Clearly, the synthetic FLAIR yields a segmentation more consistent with using real FLAIR images. Then we plot the evolution of lesions on another normal subject taken from the BLSA dataset [3]. There are 11 consecutive scans of the subject, and each of the scans is comprised of a T1w and a T2w image. Volumes of the first 10 years are registered to the 11 year by an affine registration. Then FLAIR images are synthesized for each year using another subject as atlas. The lesions are determined by applying the segmentation tool to the T1w and FLAIR images. Fig. 3 shows lesion contours of the year 1, 5 and 11. It can be seen that lesion volumes are increasing. We used the same idea on a pool of 20 normal subjects with atlas as one subject, which is not included in the cohort. Subjects who have clear lesions at baseline and have more than 7 years of scan are selected for analysis of longitudinal progression. The lesion volumes are normalized by the corresponding intracranial volume (ICV) to produce lesion fractions and are plotted in Fig. 4. CONCLUSION: We have developed a method to synthesize FLAIR images from T1w and T2w images. It is used to find longitudinal growth of lesions in the absence of an appropriate lesion distinguishing contrast. This method can potentially be seen to reduce data acquisition. In a longitudinal study, given T1w and T2w images, acquisition of FLAIR of all the subjects of all the years can be avoided, if it is acquired on a few subjects, which act as atlas. ACKNOWLEDGEMENTS: This work was supported in part by the Intramural Research Program of the NIH, National Institute on Aging and N01-AG3-2124. We are grateful to the BLSA participants and neuroimaging staff for their dedication to these studies. REFERENCES: [1] F. Rousseaui, “Brain Hallucination,” European Conf. on Comp. Vision, 2008. [2] N. Shiee et.al., “A Topology-Preserving Approach to the Segmentation of Brain Images with Multiple Sclerosis Lesions,” NeuroImage, in press. [3] S. Resnick et.al., “Longitudinal MR Imaging of Older Adults: A Shrinking Brain”, J. Neurosc., 2003. [4] B. Fischl et.al.,“Sequence-independent Segmentation of Magnetic Resonance Images,” NeuroImage, 2004. [5] A. Buades et.al., “A Non-Local Algorithm for Image Denoising”, CVPR 2005.