Impact of region-of-interest size on immune profiling using multiplex immunofluorescence tyramide signal amplification for paraffin-embedded tumor tissues

Abstract

Introduction:

Representative regions of interest (ROIs) analysis from the whole slide images (WSI) are currently being used to study immune markers by multiplex immunofluorescence (mIF) and single immunohistochemistry (IHC). However, the amount of area needed to be analyzed to be representative of the entire tumor in a WSI has not been defined.

Methods:

We labeled tumor-associated immune cells by mIF and single IHC in separate cohorts of non-small cell lung cancer (NSCLC) samples and we analyzed them as whole tumor area as well as using different number of ROIs to know how much area will be need to represent the entire tumor area.

Results:

For mIF using the InForm software and ROI of 0.33 mm² each, we observed that the cell density data from five randomly selected ROIs is enough to achieve, in 90% of our samples, more than 0.9 of Spearman correlation coefficient and for single IHC using ScanScope tool box from Aperio and ROIs of 1 mm² each, we found that the correlation value of more than 0.9 was achieve using 5 ROIs in a similar cohort. Additionally, we also observed that each cell phenotypes in mIF influence differently the correlation between the areas analyzed by the ROIs and the WSI. Tumor tissue with high intratumor epithelial and immune cells phenotype, quality and spatial distribution heterogeneity need more area analyzed to represent better the whole tumor area.

Conclusion:

We found that around 1.65–5 mm² is enough to represent the entire tumor areas in most of our NSCLC samples.

Introduction

Digital image analysis (DIA) is being used increasingly in research to accurately quantify various markers as well as coexpression of markers (cell phenotypes) within the tumor tissue to obtain reproducible information that can predict patient outcome [1].

Two different strategies based on DIA are being used to study tumor samples: (1) analyzing the whole slide image (WSI) of the tumor area and (2) analyzing representative regions of interest (ROIs) from the entire WSI. Although the WSI contains a vast amount of information and can reduce or potentially eliminate human bias [1], sometimes analysis of the WSI is impossible and inefficient because of cost, time, and resources; in such cases, selection of representative areas is an alternative approach that minimizes time [2] when we target analysis of the WSI to the tumor area only and not to other tissue structures such as normal areas and vessel structures included in a WSI. To this end, several studies have led to models that try to identify and select the correct ROIs that represent the WSI [3–5]; however, we still have not defined the minimum amount of WSI area that should be analyzed to be representative of the entire tumor.

Selecting ROIs to represent the entire tumor sometimes mimics the pathologist’s routine in analyzing samples, including scanning the entire sample in low magnification to help reveal the geographic distribution of the tumor-infiltrating lymphocytes (TILs) and to obtain an overview of the various compartments and components in the sample of tumor, such as stroma, vessels, and necrotic areas. However, the selection of the tumor is sometimes random, and most of the time we need to choose specific areas to avoid necrosis, fibrosis, or other components on the tissue that can interfere with the image analysis or can lead to false interpretations.

Selecting the number of ROIs to be representative of the tumor and choosing the location of these ROIs are critical to answering our research questions [1]. Developing consensus guideline methodologies is important to prioritizing the selection of ROIs to ensure that these regions are representative of the WSI in order to minimize interobserver variability.

In recent years, pathologists who work with DIA have encountered new challenges with the introduction of novel technologies such as multiplex immunofluorescence (mIF) tyramide signal amplification (TSA), which allow the study of simultaneous markers in a single sample [6]. If until now the study of individual markers by single-stain immunohistochemistry (IHC) was challenging, mIF introduced even more challenges for DIA. mIF staining technology has been shown to be an influential tool for immunoprofiling of formalin-fixed paraffin-embedded (FFPE) tumor samples [7–9], providing quantitative and spatial data that opens new opportunities for understanding the tumor microenvironment and enabling new therapeutic strategies [7].

The goal of this study was to define the minimum area that should be analyzed in order to represent the entire tumor in a WSI by selecting ROIs in mIF and IHC and to identify the factors that can be influenced in the selection of the areas analyzed.

Material and Methods

Image collection

All samples included in this study were surgically resected from patients diagnosed with lung cancer. The 4-µm-thick sections selected in the current study were cut from representative FFPE tumor block. Sequential sections from same cases were also prepared for H&E staining. H&E staining was reviewed by a lung cancer pathologist to evaluate fixation quality and to approve the presence of tumor as well as immune cells. Only one representative section which contain epithelial and stroma compartments with no or least necrotic tissue from each case was used for mIF study. Ten WSIs from non–small cell lung cancer (NSCLC) samples, arbitrarily selected from a previously unpublished project, included five cases of squamous cell carcinoma (SCC) and five cases of adenocarcinoma (ADC) for mIF TSA. For single IHC, 50 WSIs were randomly retrieved from another previously published project [10] of ADC (n=20) and SCC (n=30) and stained with a CD3 marker. The major subtype of ADC from this group was acinar-predominant. (online supplemental Table S1).

mIF and IHC staining and analysis

For mIF, a 4-µm-thick histologic tumor section was cut from a representative FFPE tumor block and automated stained (BOND-RX; Leica Biosystems) according to a protocol similar to one previously published [11–13] with antibodies against CK, CD3, CD8, PD-1, PD-L1, and CD68 (online supplemental Table S2). Next, the WSIs were scanned with use of the Vectra 3.0 multispectral imaging system (Akoya Biosciences) at low magnification, as previously described [11].

The total tumor area was gridded manually and included 660- × 500-µm ROIs at a resolution of 20×, 0.5-µm/pixel through the Phenochart Software image viewer 1.0.12 (Akoya Biosciences). InForm 2.4.8 image analysis software (Akoya Biosciences) was used to analyze each individual ROI and generate data by using the coexpression of markers (cell phenotypes) in the mIF panel. Six cell phenotypes were analyzed: malignant cells [total CK⁺, malignant cell–expressing PD-L1 (CK⁺PD-L1⁺)], T-cell lymphocytes [total CD3⁺, cytotoxic T cells (CD3⁺CD8⁺), antigen-experienced T cells (CD3⁺PD-1⁺)], and CD68⁺ macrophages. Hot spot ROI was defined as any ROI that had higher than the median number of CK⁺ and CD3⁺ cells than the overall ROIs; in contrast, cold spot ROI was defined as an ROI with less than the median number of CK⁺ and CD3⁺ cells than the overall ROIs. The number of individual cell phenotypes in each ROI, as well as the average of the various cell phenotypes in the whole slide, was expressed as density per mm².

For IHC, a 5-µm-thick histologic tumor section was cut from a representative FFPE tumor block and automated-stained (BOND-MAX; Leica Biosystems) against CD3. WSIs acquired with use of an Aperio Turbo 2 (Leica Biosystems) scanner at 20× magnification (online supplemental Table S2). The Image Scope toolbox from Aperio was used manually to select entire tumor areas as well as individual ROIs in 1-mm² grids (online supplemental Figure S2), which were 3 times larger than the ROI grids used by mIF (0.33 mm² each). With use of a modified nuclear algorithm tool from ImageScope (online supplemental Figure S2), we quantified all positive CD3 cells in the whole tumor area as well as in each ROI. The final number of positive cells was expressed as density by mm², equal to that determined by mIF analysis.

Statistical analysis

For mIF and IHC, automatic computational random sampling with R software was used to select 1 to 30 ROIs from each case. For each ROI (1 to 30), the sampling was repeated 10 times, and the mean density and percentage of each cell phonotype on mIF or of CD3 on IHC were calculated and associated with the total number of cells. For each sample, Spearman rank correlation coefficients were calculated between the whole tumor and the averaged value of phenotypes or CD3 from ROIs. The median and range of the correlation coefficients were also calculated. The standard deviation of correlation coefficients was calculated to assess variability. A correlation coefficient value (rho) of >0.90 was considered very strong when the number of ROIs or the percentage of each phenotype or CD3 analyzed was compared with the whole tumor. Also, the percentage of each phenotype or CD3 was calculated by dividing by the number of total cells. Density and percentage data of each phenotype were used for the correlation study. To assess the effect of individual phenotype on the number of ROIs needed to be comparable to the whole tumor, we determined the accuracy, defined as the distance from the average value of ROI to the value of whole tumor, measured as a fraction of the whole tumor measurement, and the coefficient of variation (CV), which was a measurement of the relative dispersion of data points in a data series around the mean. To evaluate the effect of intratumoral heterogeneity on the collection between ROIs and WSIs, mutant-allele tumor heterogeneity (MATH) was applied. MATH was calculated from the median absolute deviation and the median cell density of individual phenotypes [14]. The minimum areas of each sample were calculated by 0.33 µm² (mIF) or 1 mm² (IHC) × number of minimum ROIs; the percentages of minimum areas were calculated by dividing the minimum areas by the total areas.

Results

ROIs and cell markers

Figure 1A and B shows the distribution of the ROIs in one example from the ADC and SCC group by mIF; Figure 6 shows the distribution of ROIs for the same groups but determined by IHC. In mIF, the median number of ROIs that covered the entire tumor area from the WSIs was 274 (minimum 30, maximum 714 ROIs, Figure 1C, D) or 92 mm² (minimum 9.5, maximum 240 mm²) from the cases. Similarly, in IHC, the median number of ROIs that covered the entire tumor area was 96 (minimum 28, maximum 187 ROIs) or 96 mm² (minimum 28, maximum 187 mm²).

Figure 1. Representative example of a digital image analysis of an NSCLC sample using Inform software.

Each blue square represents one ROI. All ROIs that contain more than 5% no tissue areas were excluded. One high-magnification area represents a single ROI. The composite image shows the staining of all 6 phenotypes in ADC (A) and SCC (B). The total number of valid ROIs of each sample (C) and the median number of ROIs in ADC and SCC (D).

Figure 6. Sampling method in IHC.

Panoramic microphotography (4× magnification) of CD3 IHC analysis results; the entire tumor section is delimited manually with a black line, and the tumor is gridded with 1-mm squares. Squares 2 and 28 (20×) show details of positive lymphocytes for this marker.

Expression of the 6 individual immune markers was observed in most of the samples, except in one case of ADC, in which PD-1 expression was not found. By using marker coexpression, 6 cell phenotypes were studied (CK⁺, CK⁺PD-L1⁺, CD3⁺, CD3⁺CD8⁺, CD3⁺PD-1⁺, and CD68⁺) and were present in various proportions across the samples and ROIs (Figure 2 and Table 1). The median percentage of malignant cells expressing PD-L1 (CK⁺PD-L⁺) was 8.5% (range, 0.47% to 30%). When we compared PD-L1 expression in ADC and SCC samples, we found that the median percentage of CK⁺PD-L⁺ cells was slightly higher in ADC samples (11.4%; range, 2.5% to 30%) than in SCC samples (8.5%; range, 0.5% to 10.5%); however, no statistically significant differences were observed (p=0.067). Moreover, there were no statistically significant differences between ADC and SCC samples in terms of cell phenotypes densities of CK⁺, CD3⁺, CD3⁺CD8⁺, CD3⁺PD1⁺, and CD68⁺ cells. Similarly, the difference in CD3 expression, determined by IHC, between ADC and SCC was minimal.

Figure 2. Cell density (cell number/mm²) of CK⁺ tumors and TILs in all samples.

Bar graph represent the density of CK⁺ tumor cells and TILs in each sample. Density data of each phenotype are log2 transformed for better visualization.

Table 1.

Pearson’s correlation coefficient between the whole tumor area and randomly selected regions of interest or percentages of tumor area for the CD3 marker.

Marker	Tumor type	Regions of interest
		3	4	5	6	7	8	9	10	11
CD3	NSCLC	0.83	0.85	0.88	0.90	0.90	0.92	0.92	0.92	0.92
	ADC	0.71	0.80	0.80	0.84	0.86	0.88	0.88	0.88	0.91
	SCC	0.84	0.86	0.90	0.91	0.92	0.92	0.93	0.93	0.94
		Percentages of tumor area
		1%	2%	3%	5%	6%	8%	10%	15%	20%
CD3	NSCLC	0.66	0.72	0.82	0.87	0.88	0.92	0.93	0.95	0.96
	ADC	0.62	0.65	0.8	0.81	0.86	0.89	0.90	0.93	0.95
	SCC	0.69	0.75	0.83	0.89	0.89	0.92	0.94	0.96	0.97

NSCLC, non–small cell lung carcinoma; ADC, adenocarcinoma; SCC, squamous cell carcinoma.

Minimum area to represent the whole tumor area in NSCLC

We revealed that the cell density data of 1.65 mm² (from 5 randomly selected ROIs) in 90% of our NSCLC samples strongly correlated with the data in whole slides (Spearman correlation coefficient of >0.9). Overall, an average of 1.03 mm² or 2.7% of the total areas from ROIs was necessary to achieve this high correlation with WSIs in our 10 NSCLC samples (Figure 4) on mIF. Not surprisingly, in IHC, we observed that the correlation between the whole tumor and the number of ROIs analyzed increased when the number of ROIs or the percentage of tumor area increased. Overall, we observed that we needed 5 ROIs (5 mm²) to obtain a correlation of 0.9 for CD3.

Figure 4. Minimum numbers of ROIs, minimum areas, and minimum percentages of areas needed for rho >0.9.

The distribution of minimum numbers of ROIs (A), minimum areas (B), and minimum percentages of areas (C) in ADC and SCC groups. Summary of mean, median, and range data of panels A to C (D).

Histologic impact in the representation of the whole tumor area

As shown in Figure 3A and C, when we reached areas of 1.65 mm² (5 ROIs), the density data from 5 ADC samples showed that the mean correlation coefficient between ROIs and whole tumor was higher than 0.9. Among these 5 ADC samples, the area required to reach a correlation coefficient between ROIs and whole tumor higher than 0.9 was 0.66 mm² (2 ROIs). The average minimum area needed to reach rho >0.9 was 0.86 mm² (2.6 ROIs) in ADC samples (Figure 4A–D). One sample (ADC3) showed that only 0.33 mm² (1 ROI) was needed for good correlation.

Figure 3. Spearman rank correlation coefficients between ROIs and WSI.

Spearman rank correlation coefficients are calculated between whole tumor and the averaged cell density value of phenotypes from 1 to 30 ROIs. Ten sets of ROIs were randomly selected for each number of ROIs. Each dot represents the median coefficient for each sample. Colors of lines represent ADC (A) and SCC (B) tumor samples. Vertical lines represent the range of correlation coefficients. The minimum number of ROIs, minimum area, and minimum percentage of areas in each sample were needed to get the Spearman rank correlation coefficients >0.9 (C).

For SCC samples, the correlation between ROIs and whole tumor was highly sample-dependent (Figure 3B and C). Sample SCC1 had high correlation even if we chose only 1 ROI. Notably, both SCC1 and ADC3 had 2 phenotypes that possessed low positive density. The median cell densities of CK⁺PD-L1⁺ and CD3⁺PD1⁺ were 3 (range, 0 to 437) and 9 (range, 0 to 1068) in SCC1, respectively. The median cell densities of CD3⁺CD8⁺ and CD3⁺PD1⁺ were 27 (range, 0 to 173) and 0 in ADC3, respectively. In contrast, sample SCC4 had lower correlation, with the correlation coefficient value first reaching 0.9 at 2.64 mm² (8 ROIs). Interestingly, when the numbers of ROIs increased to 9.9 mm² (30 ROIs), we found that almost half of the correlation coefficient values decreased to less than 0.9. Moreover, possibly due to the high heterogeneity of the tumor, the ranges of absolute values of the cell densities of 6 phenotypes in each ROI were high, and the outliers of cell density values of each phenotype may have affected the correlation. However, all of the values were greater than 0.8, which suggested a strong correlation with the WSI. Therefore, we chose 2.64 mm² as the minimum area for sample SCC4. The average minimum areas needed to reach rho >0.9 were 0.86 mm² (2.6 ROIs) in ADC and 1.19 mm² (3.6 ROIs) in SCC samples (Figure 4A–D). Therefore, we transformed the absolute number to percentage data of each phenotype by dividing the number of total cells on each phenotype. As shown in online supplemental Figure S3, when the areas are smaller than 1.98 mm² (6 ROIs), the percentage data from 5 ADC samples presented slightly better correlation between ROIs and the whole tumor. Remarkably, when the areas reached 0.99 mm² (3 ROIs) in SCC samples (except SCC4), the correlation coefficients between ROIs and the whole tumor were higher than 0.9.

On the basis of CD3 expression, we determined on IHC that the total number of ROIs evaluated in each case depended on the size of the total tumor area. We excluded two cases from further analysis that had 0 or 1 ROI; thus, 29 SCCs and 19 ADCs were included in the final statistical analysis. The results of DIA showed that CD3 densities were slightly lower in SCC (mean, 1285 cells/mm²) than in ADC (median, 1687 cells/mm², P=0.062), but they had a similar dispersion grade. In both groups, we considered atypical samples to be those that had a higher density of positive cells than did standard samples (> ~3 cells/mm²). Considering the number of ROIs and the percentages of tumor area analyzed, we observed a tendency toward positive lineal regression for CD3 and tumor histologic type, as shown in Table 1.

In SCC IHC, we found correlations of 0.84 with 3 ROIs and 0.69 and 1% tumor area. In ADC IHC, we found correlations of 0.71 with 3 ROIs and 0.62 with 1% tumor area. The correlation was 0.96 with 20 ROIs and 20% tumor area. In SCC, we observed a high and stable correlation of 0.90 with 5 ROIs and 8% tumor area for CD3. Similar results for 20 ROIs and 20% tumor area were found for ADC (0.94 and 0.95, respectively, for CD3). Furthermore, we observed a correlation of 0.84 for CD3 with 6 ROIs; a correlation of 0.90 was also found for CD3, with 8% observed in ADC (Figure 7A and B).

Figure 7. CD3 correlations.

Correlation between CD3 and region of interest (A) or the percentage of tumor area (B). The green line indicates a correlation of >0.90.

Influence of individual phenotypes or markers in the ROIs analyzed

Next, we assessed the influence of individual phenotype on the minimum areas needed for rho >0.9. We used accuracy to evaluate the calculated results of ROIs’ similarity to WSI results. We used the CV to compare the distribution of individual phenotype of ROIs to that of WSIs. Online supplemental Figure S4 clearly shows patient influence on accuracy and CV. With increased areas, the spread of accuracy narrows. Therefore, we concluded that both patient and phenotype influence correlation between ROIs and WSIs. Moreover, although intra tumor heterogeneity (ITH) is an important factor for each sample, it may not have a monotonic relationship with correlation between the whole tumor and ROI data in our dataset.

Influence of hot spot ROIs on overall choice to represent the WSI

To see the influence of hot spot ROIs, we defined the hot spot with tumor cell marker CK and TIL marker CD3. Hot spot ROIs contained both CK⁺ cells and CD3⁺ TILs with cell density values that were higher than the median values. In contrast, cold spot ROIs contained cell density values that were less than the median values of CK⁺ cells and CD3⁺ TILs. Hybrid spot ROIs had CK⁺ cell and CD3⁺ TIL density values that were higher than the median values in one and less than the median in the other. We only tested areas up to 3.3 mm² (10 ROIs) due to the reduced number of ROIs in the subgroups. As shown in Figure 5A and B and online supplemental Figure S5, we observed better correlation to WSIs with hot spot ROIs than with cold spot or hybrid spot ROIs in both ADC and SCC samples (except SCC4). Only 0.66 mm² of hot spot area (2 hot spot ROIs) showed the correlation coefficients to whole tumor were higher than 0.9 in most of the samples.

Figure 5. Spearman rank correlation coefficients between hot spot ROIs and WSI.

Spearman rank correlation coefficients were calculated between the whole tumor and the averaged cell density value of phenotypes from 1 to 10 hot spot ROIs. Ten sets of ROIs were randomly selected for each number of ROIs. Each dot represented the mean coefficient for each sample. Line colors represent ADC (A) and SCC (B) tumor samples. Vertical lines represent the standard deviation of correlation coefficients.

Finally, we calculated the minimum percentage of area that was needed in order to get good correlation to the whole tumor. The average percentage of area was 1.89% (median, 0.60%; range, 0.33% to 6.86%) in ADC samples. In SCC samples, the value was 3.48% (median, 1.12%; range, 0.24% to 13.85%) (Figure 4D). For all NSCLC samples, average minimum areas of 2.7% were needed for ROIs to have very strong correlation with WSIs.

Discussion

In the current study, we aimed to define the minimum area that can represent the entire tumor in a WSI in mIF and IHC images. For mIF images, we found that a cell density from 2.7% of total areas (5 randomly selected ROIs, about 1.65 mm²) was enough to achieve a strong correlation of more than 0.9 (Spearman correlation coefficient) in 90% of our samples; for single IHC images, we revealed that a correlation value of more than 0.6 was achieved by using 1% of total areas. Interestingly, we observed that, in IHC images, about 5 mm² was enough to represent the entire tumor areas in most of our NSCLC samples. We reasoned that larger ROI might contain more variation in IHC leading bigger areas was required for Spearman correlation coefficient to reach 0.9.

Automatic or manual random sampling to select areas with high- and low-density markers is commonly used for large-scale tumor studies, most of which use an optical microscope. However, it is crucial that a sufficiently large area is sampled to obtain an accurate representation of the whole tumor and thus avoid errors in the interpretation and final results [3, 4, 15]. A great variety of tissue sampling methods exist, and in DIA, there is no standard approach to determining the amount of tissue that should be analyzed to provide a good representation of the tumor, causing variability in the results.

Each marker and tumor type need to be considered when selecting the appropriate number of ROIs for analysis. Representative images of tumors are needed in research and clinical practice to allow effective evaluation and to obtain accurate results, saving time and resources.

mIF is a vital tool for immunoprofiling of tumor samples [16–20], and typically, the investigation process requires the selection of ROIs within the sample for analysis. Tumor heterogeneity and sampling bias usually result in data that may not be representative of the whole tumor. The present study compared immunoprofiling data from ROIs with whole tumor analyses in NSCLC with use of mIF image analysis.

Our mIF results revealed that a small number of ROI analyses provide acceptable correlation with data from the WSI. In ADC, data from 3 ROIs had very strong correction with the WSI (rho >0.9). Four ROIs were needed in SCC samples in order to get rho >0.9 with the WSI. Although NSCLC showed high variability, especially for a particular TIL population, the random selection of 5 ROIs in 90% of our samples generated data that vastly correlated with the WSI. These results suggest that investigators should pay special attention to the image analysis strategy when high tumor heterogeneity is present.

We had 6 phenotypes in our mIF study, with a wide range of cell densities in individual phenotypes. By analyzing the accuracy and CV, we found that every phenotype has an important role in the correlation between ROIs and WSIs. Additionally, ITH may not have a monotonic relationship with the correlation between ROIs and WSI results in our dataset.

TIL distribution usually shows heterogeneity across the dimensions of a single section or regions of a given tumor. Single marker staining by chromogenic IHC defines the hot spot as the area with the highest density of immunoreactive cells within the tumor. Several studies found that hot spot enumeration is equivalent to whole tumor enumeration for prognostication in breast cancer [21]. For mIF analysis, cell densities of both tumor cells and TILs are critical for the study. We defined the hot spot in this study as any ROI that contains a high number of CK⁺ cells (> median) and high number of TILs (> median). Seven of ten samples showed that having 2 hot spot ROIs is good enough to get rho >0.9 with WSI. The reason that SCC4 was an exception may have been due to the high total number of ROIs and low percentage of hot spot ROIs in the whole tumor area (13.9%). These findings suggest that, when selecting the ROIs for mIF analysis, the distribution of both CK⁺ cells and TILs should be equally considered, whereas hybrid ROIs should be avoided.

We also identified the minimal amount of tissue required to characterize entire NSCLC tumors stained with chromogenic IHC. Our DIA approach showed that results varied by marker and histologic type, but in all cases, 5 randomly selected ROIs (1 mm² each) or 20% tumor area resulted in a correlation of >0.90.

The entire tumor is not commonly sampled in NSCLC; 3 or 5 ROIs at 20× or 40× magnification are frequently used [22–25] and in most cases, the number depends on the pathologist criteria, which could lead to interobserver variability if the study is reproduced. The grid method is another method of sampling tissue; however, the optimal size of the grid and the number of areas to be analyzed have not been standardized [26]. In our study, we observed a minimum ROI (at 20× magnification) as representative of the tumor and determined that the percentage of tumor area needed for analysis directly correlated with the type of marker being analyzed; this finding is in concordance with the results of a study by Johnson and colleagues,[22] who reported that densely and diffusely distributed markers, such as CD3, require a smaller representative area for analysis than do those sparsely and patchily distributed, such as FOXP3. Finally, tumor type needs to be considered when sampling tissue in NSCLC.

This study was limited to NSCLC. Further studies are needed to validate our findings and to establish a pathology reference for image analysis that can integrate the details of the various aspects as histology, tissue size, and markers used to potentialize a best approach for image analysis.

Abbreviations:

ADC

adenocarcinoma

CV

coefficient of variation

DIA

digital image analysis

IHC

immunohistochemistry

mIF

multiplex immunofluorescence

NSCLC

non–small cell lung cancer

ROIs

regions of interest

SCC

squamous cell carcinoma

TIL

tumor-infiltrating lymphocyte

TSA

Tyramide Signal Amplification

WSI

whole slide image

Data Availability Statement:

All data generated or analyzed during this study are included in this article and its supplementary material files. Further enquiries can be directed to the corresponding author.