• No results found

Identifying distinct regions in a MIMS image is an important preprocessing step both for subsequent visualization and data analysis. As shown in Fig. 4·4, each MIMS image includes three distinct periods: an initial laser off blank period (blue), a laser

on period (red), and a laser off washout period (green). Additionally, when imaging

a sample, the imaged area during the laser on period can be segmented into two regions: data associated with the sample and a background region. Segmenting these data is particularly challenging because the properties associated with each of these distinct regions vary depending on the isotope under consideration and the material being ablated. Generally, the signal during the blank and washout periods tends to be the lowest while the signal within the laser on portion is higher. A variety of factors including isotope sensitivity, sensor drift, polyatomic interferences, material properties, etc., result in images where this general pattern may not hold.

Before defining the segmentation procedure in detail, notation that will be used throughout this chapter is needed. Figure 4·9 provides a reference to the notation used. Consider a MIMS sequence, −→I , with T -types of samples and d-channels ac- quired. Each data type, τ = 1, . . . , T , in the MIMS sequences can be thought of as a d-channel, vector-pixel image, −→I τ. The i-th channel of the image

− →

I τ will be

indicated as Iτ(i) where i = 1, . . . , d, and the pixel values in each channel indicate the intensity in counts per second (cps) detected by the MS for a distinct isotope. For example, i = 1 may represent the metallomic intensity for13C, i = 2 the intensity for

(a) Custom pseudo-coloring (b) Multi-channel visu- alization

(c) Graphical user interface for manual affine registration

Figure 4·8: Examples of select image processing and rendering tools. (a) Depiction of a calibrated MIMS image with a fire colormap with

ppm values indicated by the colorbar. (b) Multi-channel psuedo-colored

image depicting four channels of data 56F e [red], 63Cu [blue], 64Zn

[green],56Gd [white] (c) Manual affine image registration GUI where I

M

(moving image) is displayed as edges overlayed on the IF (fixed image).

The mutual information value is given as well as the six values in the transformation matrix A, discussed in more detail in Section 4.5.2.

140Ce, i = 3 the intensity for 65Cu, and so on.

The objective of MIMS segmentation is to assign one of four labels to each pixel in the image−→I τ: (1) blank, (2) washout, (3) background, and (4) sample. Notice that for the

MIMS sequence shown in Fig. 4·9, the 13C and 66Zn channels for the “brainHIPP”

image,−→I7, have vastly different intensity properties for each of the regions of interest.

Because the sequence of events on each line follows the sequence of events detailed in Fig. 4·4, channels which have properties consistent with the labels of interest can be identified. The function leverages the intensity characteristics of each channel to ultimately create a mask Mτ ⊆ {1, 2, 3, 4} whose values correspond to the label of

each pixel in the image.

In this thesis, the MIMS sample types can be categorized into three distinct groups: certified reference materials, calibration standards, and samples of interest. Samples of interest are primarily biological samples including sections of mouse brains (Chap- ter 5) and human lacrimal sacs (Chapter 6), but for optimization of the acquisition procedure and instrument settings, 3.05 mm copper grids (Electron Microscopy Sci- ences, 2016) were also imaged. The primary certified reference material used was the NIST glass 612 (National Institute of Standards and Technology, 2016) and the cal- ibration standards were prepared in our laboratory. Both the NIST and Calibration images have masks Mτ ⊆ {1, 2, 3} since the laser on portion of the image does not

have two distinct regions as MIMS of samples do.

Ideally, the intensity properties of the different regions of the image would have sep- arable modes, making image segmentation by k-means clustering or identification of Gaussian components (by parameterization of a Gaussian mixture model (Bishop, 2006)) straightforward approaches. While it may be true that the intensity char- acteristics of some channels in a MIMS image exhibit a high degree of separability,

this is not always the case and automated identification of such channels would be required nonetheless.

The segmentation algorithm uses a rule-based approach to identify channels whose intensity characteristics clearly represent either the laser on region, the sample region, or neither. The algorithm is dependent upon the line characteristics detailed in Fig. 4·4. The ground-truth mask is expected to have the following properties:

1. A minimum blank period of 5 sec

2. A sample which takes up at least 25% of the total imaging area and is not represented in the blank region of the image

For each channel in the image, a bimodal segmentation algorithm identifies the two most prominent histogram peaks then uses the minimal intensity between these peaks as a threshold to produce an intensity mask Kτ(i) ⊆ {0, 1}. The properties of the mask

K(i)

τ are evaluated and the mask is categorized as either being representative of laser

on area (Mτ > 2), of sample area (Mτ = 4), or of neither. Figure 4·10 shows an

example of a channel image Iτ(i) along with the mask Kτ(i) for each of the possible categories. Once all d channels are evaluated, the channel masks in each category (g = {laser on, sample}) are used to create a segmentation image defined as:

Sτg(p) =X

i∈c

Kτ(i)(p) ⊆ {0, 1, . . . , |i ∈ g|} Two cluster k-means is then applied to Slaser on

τ (Fig. 4·11(a)) to generate a k-means

label image Kτlaser on ∈ {0, 1} (Fig. 4·11(b)). The k-means result is checked to ensure the laser on part of the image is assigned a cluster label of one and is filled. The output is then used to create a preliminary mask of the MIMS type Mτ0 (Fig. 4·11(c))

defined as follows:

Mτ0(p ∈ first half of columns) =

         1 Klaser on τ (p) 6= 1 3 otherwise

Mτ0 (p ∈ second half of columns) =

         2 Kτlaser on(p) 6= 1 3 otherwise

Note that Mτ0 = Mτ for the NIST and calibration standard MIMS types. For sample

types, the part of Mτ0 = 3 will be further divided into background and sample using the channel masks categorized as sample masks. Again, using Ssample

τ (Fig. 4·12(a)),

a two cluster k-means mask is generated (Fig. 4·12(b)), ensuring the background part of the image has a cluster label of zero. With Mτ = Mτ0 as a starting point, the

sample pixels are then defined as follows:

Mτ(p) =          4 Mτ0(p) = 3, Kτsample(p) = 1 Mτ0(p) otherwise

(a) Example of laser on channel image and mask

(b) Example of sample channel image and mask

(c) Example of neither channel image and mask

Figure 4·10: Examples of the channel categorization routine used by the segmentation algorithm. The top row depicts the intensity adjusted channel images I(i)

τ and the bottom row depicts the associated mask

K(i)

τ . Column (a) shows a representative laser on channel, column

(b) shows a representative sample channel, and column (c) shows an example which does not get assigned a category.

(a) Slaser on

τ (b) Kτlaser on (c) Mτ0

(a) Ssample

τ (b) Kτsample (c) Mτ

Figure 4·12: Intermediate images used to define the final MIMS seg- mentation mask

Related documents