Biomedical Engineering Reference
In-Depth Information
14.3.4
Analysis of Brain Anatomy in Multiple Sclerosis
Besides applications which are primarily research-oriented, automatic anal-
ysis of medical image data is becoming increasingly important in the evalu-
ation of drug therapies. Especially in phase II
III clinical trials, accurate and
reproducible quantification of disease-specific features in MRI data is key to
the acceptance of new treatments.
For example, the quantification of total lesion load (TLL), sometimes
referred to as “burden of illness,” in T2-weighted MRI scans of patients with
multiple sclerosis (MS) has become a standard surrogate endpoint in clinical
trials. To date, a modified version of the INSECT pipeline has been used to
quantify TLL in two large-scale, multicenter clinical trials in MS: (i) phase III,
600 patients, scanned 3 times on a yearly basis using T1-, T2-, and PD-weighted
MRI; and (ii) phase II, 150 patients, scanned 6 times on a monthly basis using
T1-, T2-, PD-weighted, and FLAIR MRI. The MRI acquisiton protocol used in
these studies was uniform accross centers and the resulting image data were
carefully controlled for quality.
The version of INSECT used for the quantification of MS lesion volume dif-
fers somewhat from the one used for the study of normal brain. First, a stan-
dard stereotaxic brain mask is used to eliminate false positive lesions
possibly detected outside the brain area, and stereotaxic SPAMs of WM, GM,
and CSF are included as extra features in the classification process to sup-
press false positive lesions inside the brain based on their (stereotaxic) loca-
tion. Second, the MRI data volumes are all intensity-normalized to a standard
brain model in stereotaxic space, allowing for the use of a once-trained, fixed
classifier across all acquisitions in the study. The classifier used is an artificial
neural network, trained to separate MS lesion from other tissues on a limited
number of hand-labeled volumes. The resulting output is a binary map of MS
lesions for each patient scan. The accuracy of INSECT-obtained MS lesion
measurements has been validated against those obtained manually by
experts.
98
Figure 14.11 shows the MS lesion SPAM generated from a total of
460 automatically segmented patient data sets.
FIGURE 14.11
Average MS T1-weighted data and lesion SPAM (
460), shown in transverse cross-section
and rendered with the CSF SPAM shown in Figure 14.5.
n
