4.5 Multi-modal Image Registration
4.5.6 Discussion
Through these experiments, the functionality of the proposed multi-channel, multi- modal image registration function was established. A number of factors influencing registration accuracy and general properties of the image registration function were observed. One such observation was the effect of cost function smoothness on regis- tration performance. The plots in Fig. 4·28 show the projection of the MI at each iteration of SA onto each of the transformation parameter axes. It is apparent that compared to Fig. 4·28(c), the distribution of the data in Figs. 4·28(a) and 4·28(b) have peaks that are more clearly defined for nearly all of the parameters. These sharp peaks are typically a reliable indicator of a “good” image registration result.
A second interesting and motivating reason for estimating multi-dimensional MI rather than other perhaps more simple schemes (such as maximization of the sum of MIs) reveals itself in Fig. 4·29. Figures 4·29(a) and 4·29(b) show MI values re- sulting from registration between two different fixed scalar images. Figure 4·29(c) shows data from an optimization function which maximized the sum of MI between two scalar images for each of the channels. In Fig. 4·29(c), the scaling peaks from both the 56F e image and the 63Cu image are present, resulting in an ambiguous
or non-specific value of sy which maximizes MI. However, when computing MI for
vector-valued images, it results in a singular peak in the parameter space, as seen in Fig. 4·29(d). While the MI in Figs. 4·29(a) and 4·29(b) is peaked in the sy parameter
space, this figure shows that final values reached using different channels are vastly different (exemplary final registration results can be seen in Figs. 4·25(b) and 4·25(c)). Additional take-away points were the improvements in computational time without a significant effect on registration accuracy when using the Epanechnikov kernel rather than the Gaussian (Normal) kernel to compute the joint probability function. While
(a) Exp. 1 data (b) Exp. 2 data
(c) Exp. 3 data
Figure 4·28: Examples of cost function projections onto transforma- tion parameter axes for each of the MMMCIR experiments.
the results shown in this work are largely qualitative, the algorithmic infrastructure needed to perform multi-modal, multi-channel image registration has been developed and validated.
4.6
Summary of Contributions
The focus of this chapter has been on the algorithmic methods developed to address the technical challenges associated with visualization, analysis, and interpretation of MIMS data. Because the field of metallomic imaging mass spectrometry is still in it’s early stages, standard methods have yet to be defined. Contributions made towards addressing the challenges associated with MIMS data can be summarized as follows:
Figure 4·29: The data are from Exp. 3 image pairs. The images show various results for MI projected onto the sy parameter axis for
• Development of functions for MIMS raw data extraction, segmentation, and visualization (Sections 4.1 and 4.2)
• Development of methods for representations of MIMS in a standard domain using calibration standards or image intensity properties (Sections 4.3 and 4.4) • Development and validation of a novel method for multi-modal and multi-
channel image registration using mutual information (Section 4.5)
These efforts have been used to provide evidence to support hypothesis of the un- derlying physiological events occurring in mouse models of traumatic brain injury (Tagge et al., 2017). Additionally, MIMS data extraction and visualization have been fundamental in the advancement of the data acquisition process, facilitating the establishment of this imaging method for precise analytical analysis of biological samples.
Chapter 5
Analysis of Metallomic Brain Images of
Nanoparticle-Injected Mice
5.1
Experimental Objective
Metallomic imaging mass spectrometry (MIMS) is a method uniquely posed to pro- vide quantitative information about the elemental-isotopic composition and spatial localization of a biological sample. As described in Section 1.3, elements which do not naturally occur in the brain may be intravenously injected to probe the degree of blood-brain barrier (BBB) dysfunction. The cocktail may be injected at various stages of the experiment depending on the goal of the assessment. In addition to in- jection of exogenous compounds, evaluation of physiologically-relevant elements and isotopes may also provide useful insight into the mechanisms of injury and the effects on the various brain regions.
Results from a case study of mice injected with a gadolinium-based MRI contrast agent suggest that mice exposed to an impact neurotrauma (INT) on the left-side of the head sustain left-sided disruption of the blood-brain barrier (BBB). As described in a recently submitted publication based on these results, and shown in Fig. 5·1 (Tagge et al., 2017):
“Focal BBB disruption and colocalizing serum albumin extravasation detected in the brains of living mice by dynamic contrast-enhanced MRI (DCE-MRI) neuroimaging with gadofosveset trisodium, an FDA-approved gadolinium-based contrast agent that binds serum albumin. High-field (11.7T) T1-weighted MRI (T1W-MRI, A and B) and DCE-MRI (C, D) with systemically administered gadofosveset trisodium. T1W-MRI and DCE-MRI were conducted 3 hours (A, T1W-MRI; C, DCE-MRI) and 3 days (B, T1W-MRI; D, DCE-MRI) after im- pact (IMP) or control (CON) exposure. T1W hyperintensity (A, B) colocalized with BBB permeability defect detected by DCE-MRI (C, D) in the left perirhi- nal cortex (arrows) 3 hours and 3 days after IMP but not CON exposure. Nonspecific signal was observed in the ventricles and sagittal sinus.”
The157Gd MIMS brain image for the impact exposed mouse (Fig. 5·1F) also exhibits
a clear hyperintensity of the gadolinium signal colocalized with the in-vivo detected areas of BBB dysfunction.
Using the tools described in Chapter 4, MIMS brain images were registered with images from the Allen Mouse Brain Atlas (AMBA) (Allen Brain Atlas, 2013; Lein et al., 2007) to automate evaluation of signal levels in different subregions of the brain. Regions on the left and right side of the brain are used to explore whether individual mice in the INT cohort exhibited differences in elemental gadolinium distribution indicative of BBB disruption.
The goals of this Chapter are two-fold:
1. To demonstrate the use of the multi-modal, multi-channel image registration (MMMCIR) function to impart anatomical boundaries on MIMS brain image, and
Figure 5·1: 157Gd MIMS brain image for BBB compromise proof of
concept. Full explanation for A–D is provided in the text. (A) T1W- MRI at 3 hours post injury. (B) T1W-MRI at 3 days post injury. (C) DCE-MRI at 3 hours post injury. (D) DCE-MRI at 3 days post injury. (E) 157Gd MIMS brain image from a control mouse. (F) 157Gd
MIMS brain image from an impact (IMP) exposed mouse. Anatomical compass legend: D, V; dorsal, ventral. L, R; left, right. (copied from (Tagge et al., 2017, Fig.5))
2. To evaluate the signal levels in counts-per-second of left- versus right-sides of the brain in a cohort of MIMS brain image from approximately the same coronal location.
This chapter will first present the specifics of the tissue preparation and image selec- tion in Section 5.2. Section 5.3 will detail the image preparation and the processing procedure done to extract the results shown in Section 5.4. A comparison of various experimental design choices will also be analyzed in Section 5.4.