4.5 Classification of CD4+ T cells, B cells, and Monocytes
4.5.2 Classification of cell subsets with the use of DHM
Phase maps were generated for each cell from the recorded interference pattern, following the procedure detailed in section 4.4.4. An example phase map for each cell type can be seen in figure 4.11 alongside its respective white light image. The white light images illustrate that monocytes are quite distinct and are much larger in size than T- or B-cells. T cells and B cells however pose a particular challenge as they are very similar in terms of morphology. This is often the limiting factor for identification by techniques such as flow cytometry, which relies on differences in size and shape to generate different scattering profiles.
Radi
ans
FIGURE4.11: Top row represents a typical white light image for each of CD4+ T
cell, B cell, and monocyte cell types. The scale bar denotes5µm. The bottom row shows the respective phase map for each cell. The intensity shows the phase dif- ference between the object and reference waves, which is directly proportional to optical path difference. Colour bar represents phase differences in units of radians.
The phase maps provide a measure of the OPD between the reference and object waves (equation 4.7) and can therefore be considered a map of the optical thickness across the cell. Optical thickness is related to both the absolute thickness of the cell and its intracellular structure. The intracellular structure inherently causes varia- tions in the refractive index through the cell. Optical path length (OPL) can there- fore be calculated as the integral of the absolute thickness, s and refractive index, n variations along a path, p, as shown in equation 4.8.
OP L= Z
p
n(s)ds (4.8)
To separate sample thickness from refractive index is non-trivial and would re- quire a more sophisticated set-up, such as dual wavelength illumination [248, 249], scanning of the illumination angle [250, 251], or the introduction of phase shifts [252]. For the studies presented here measurement of the optical path length is sufficient and knowledge of the absolute cell thickness was not necessary.
For subsequent analysis the phase maps were described either in terms of a his- togram of the pixel intensity values or by four texture parameters. First, a histogram was generated for each phase map, representing the number of pixels within specific intensity value ranges. A typical histogram for each of the three cell populations is shown in figure 4.12. Information may be extracted directly from these histograms regarding the cell size (the total number of non-zero pixels is equivalent to the 2 dimensional area), maximum OPD (largest pixel intensity value), and the phase vol- ume (total of all pixel values). With each cell represented by a histogram vector a training data set was formed on which PCA was conducted. It should be noted that PCA may also be able to recognize patterns such as how evenly spread the values are, which is indicative of the uniformity across the phase map.
Scatter plots were produced using the first 3 PCs and can be seen in figure 4.13 A-C. PC1 shows an excellent ability to identify monocytes, which are morphologi- cally quite different from T- and B-cells. Monocytes are typically larger in size and thicker, resulting in a larger total OPD and a higher maximum OPD value. This is consistent with the observations from both white light images (figure 4.11) and the histogram plots (figure 4.12). Higher order PCs are required to recognise more sub- tle morphological differences. It is PC3 that has the ability discriminate between the morphologically similar T cells and B cells, as observed in figure 4.13 B-C.
LOOCV statistics were applied to the full data set to quantify the discrimination ability of using histograms to represent the DHM images. Excellent sensitivity and specificity values were obtained and are summarised in table 4.4. Sensitivity and specificity values of 86.8% and 98.6% respectively were achieved for CD4+ T cells
N u mb er o f Pix els N u mb er o f Pix els N u mb er o f P ix els
Optical Path Difference (a.u.)
Optical Path Difference (a.u.)
Optical Path Difference (a.u.)
FIGURE4.12: A typical histogram generated from the phase map of a B cell, CD4+
T cell, and monocyte (top to bottom). The range of pixel intensity values are recorded, which are related to the optical path lengths across a cell.
and B cells, 98.7% and 100% respectively for CD4+ T cells and monocytes, and fi- nally 100% and 100% respectively for B cells and monocytes. The morphological similarity of T- and B-cells is reflected in the lower sensitivity and specificity values obtained in contrast to comparisons with the morphologically distinct monocytes, where sensitivity and specificity values of up to 100% are achieved.
The observed phase difference between B- and T-cells may be related to differ- ences in their intracellular composition. Previous studies have investigated morpho- logical differences between B and T lymphocytes finding variations in the amount of cytoplasm, stippled chromatin, and nuclear morphology; such as nuclear size, homogeneity, nuclear folds, thickness of nuclear membrane, and presence or uni- formity of nucleoli [253, 254]. The most distinctive features reported by Parker et al were in the nuclear morphology, where B cells frequently have a characteristi- cally round and regular nucleus with thick nuclear membrane, in contrast to T cells which typically have a nucleus with deep folds and an irregular configuration, stip- pled chromatin, and small or absent nucleoli. Additionally B cells often have an
P C 2 P C 2 P C 3 P C 3 P C 3 P C 3 PC1 PC1 PC1 PC1 PC2 PC2 Monocyte Monocyte CD4 + CD4 + B cell B cell A) E) F) D) C) B) Analy sis b y h is togr am Analy sis b y TA
FIGURE 4.13: Scatter plots using the first three PCs for discrimination between
CD4+ T cells, B cells, and monocytes. The top row (A-C) illustrates the discrimina- tion ability when analysing DHM phase maps by means of a histogram descriptor vector. The bottom row (D-F) illustrates the discrimination ability when analysing DHM phase maps by means of four texture parameters. When employing the his- togram vector method compared to TA, clusters appear to be more distinct and
show less overlap between T- and B-cell populations.
abundant cytoplasm in comparison to T cells [253].
A second method of analysis was investigated, employing TA to characterise the phase maps. The four texture parameters calculated were contrast, correlation, energy, and homogeneity. The average values, and respective standard deviations, for each of these parameters for each cell type are recorded in table 4.3. Contrast and energy appear to be the most useful parameters, showing the most variation, particularly between monocytes and B cells or CD4+ T cells.
A vector made up of these 4 parameter values was generated for each phase map and PCA was conducted on the new training data set. Scatter plots were gen- erated using the first 3 PCs (figure 4.13 D-F). Observing the PC loadings revealed that contrast was the most significant parameter and contributed most to PC1, and that energy contributed most to PC2. It can be observed in figure 4.13 D and F that PC2 (relating to energy) plays an important role in discriminating monocytes from B cells or T cells. The clusters correlating to B- or T-cells however show significant
Contrast Correlation Energy Homogeneity CD4+ average 0.0388 0.9971 0.6718 0.9853 std. dev. 0.0091 0.0006 0.0294 0.0012 B cell average 0.0382 0.9961 0.6619 0.9828 std. dev. 0.0082 0.0015 0.0637 0.0032 Monocyte average 0.0803 0.9963 0.4121 0.9692 std. dev. 0.0161 0.0008 0.0347 0.0043
TABLE4.3: Average texture parameters for each cell subset. The most useful pa- rameters for discrimination appear to be contrast and energy.
overlap and would be very difficult to successfully discriminate between.
PCA and LOOCV was applied to the whole data set and pairwise sensitivity and specificity values were calculated, which are summarised in table 4.4. Sensitivity and specificity values of 78% and 62.2% respectively were achieved for CD4+ T cells and B cells, 98.5% and 97% respectively for CD4+ T cells and monocytes, and finally 100% and 100% respectively for B cells and monocytes. The lower values achieved between CD4+ T cells and B cells is expected as they are morphologically quite sim- ilar to each other and their respective clusters in PC space were not well defined. In this sense analysis by TA does not show as strong a discrimination ability as when the phase maps were characterised by histograms.
Raman DHM DHM
Spectroscopy Histogram Texture analysis
sens spec sens spec sens spec
% % % % % %
CD4+ v B cell 86.8 98.6 93.8 85.4 78.0 62.2
CD4+ v Monocytes 97.9 98.1 98.7 100 98.5 97.0
B cell v Monocytes 98.6 98.1 100 100 100 100
Average 94.4 98.3 97.5 95.1 92.2 86.4
TABLE 4.4: Summary of sensitivity and specificity values achieved for each method of analysis. Raman spectroscopy and DHM are each capable of efficiently discriminating between cell subsets. Phase maps were either described in terms of a pixel intensity histogram or by four texture parameters; the use of histograms
proved to be more efficient than using texture analysis.
TA was further investigated by calculating the four texture parameters along four directions (0◦, 45◦, 90◦, 135◦), producing a vector of 16 values for each phase
map. However this did not show any improvement in the discrimination ability. TA along different directions is commonly employed for analysing tissue samples where directional features may be expected. However this is not an expected characteristic when analysing whole cells, particularly as they are at random orientations on the quartz slide.
Finally the two vectors (histogram and TA) were concatenated to generate a new descriptor that contained both OPD and TA information. PCA and LOOCV were applied to this data set and did not show any significant improvement in comparison to analysis by the histogram vector alone. It can be concluded that a histogram of the pixel intensity values is the most efficient method of characterising a phase map for discrimination between the three cell subsets.