Quantitative image analysis for CD8 score in lung small biopsies and cytology cell-blocks

Abstract

Introduction:

Immunotherapy has shown promising results in non-small cell lung cancer (NSCLC), for which tumour-infiltrating cytotoxic (CD8+) T cells play a critical role. We investigated the utility of image analysis (IA) to quantify CD8+ T cells in a series of matched small biopsies and resections of NSCLC.

Methods:

CD8 immunohistochemistry was performed on cell-blocks (CB), core needle biopsies (CNB) and corresponding resections from primary NSCLCs. Slides were digitised using an Aperio AT2 scanner (Leica) and annotated by whole slide image (WSI) or fields of view occupied by tissue spots (TS). Quantitative IA was performed with a customised Aperio algorithm (Leica). CD8 scores (number of T cells with 1–3+ staining/total area) were then compared.

Results:

Forty-four cases with CB or CNB material and a corresponding resection were analysed. Average CD8 score was determined in CB (7.67 WSI, 77.67 TS) and/or CNB (47.35 WSI, 325.67 TS), and corresponding resections (190.35 WSI, 336.58 TS). CD8 score concordance was highest (78.6%) for CNBs using WSI annotation. Overall, small biopsies (CB or CNB) correlated with the resection in 71.4% cases using WSI and 63.3% cases using TS annotation. IA performed better for low CD8 scores.

Conclusions:

These findings show that CD8 density in NSCLC can be quantified by IA in small biopsies and cell blocks, achieving the best concordance using WSI scores. Discrepancies were attributed to values near the cut-off and background detection of staining. These data warrant future studies with more cases and follow-up data to further investigate the clinical utility of IA for CD8 analysis in NSCLC.

1 |. INTRODUCTION

The tumour microenvironment (TME) has been a major focus of cancer research for theranostic purposes, given that many studies have shown the importance of the immune system in the host defence against malignancies and the benefit that cancer patients can obtain from immune checkpoint inhibitors.^1–3 In particular, the infiltration by CD8-positive (CD8+) cytotoxic T-lymphocytes in the TME can be predictive of survival in various tumours.^4–9 In non-small cell lung cancer (NSCLC), immunotherapy has shown promising results and tumour-infiltrating T cells play a critical role in this process, particularly CD8+ T cells.^7–13 In addition, the density of different intratumoural immune cells (eg, dendritic cells, macrophages, and CD8+ T cells) has been shown to correlate with various clinicopathological factors including smoking status, pathological stage and different molecular alterations.^14,15 Furthermore, there is an increasing awareness that traditional tumour-node-metastasis (TNM) staging provides incomplete prognostication, and that molecular findings and characterisation of the TME can provide an enhanced ability to predict outcomes in NSCLC patients, particularly those that may be in the same TNM staging group.⁷ The importance of immunoscoring has accordingly become more mainstream, particularly in colorectal cancer where the idea of an immunoscore, combining analysis of the density and type of tumour-infiltrating lymphocytes within different tumour and stromal compartments, can now complement the TNM staging for colorectal cancer.^16,17 These findings have led to an increased demand for pathological evaluation of the TME on tumour specimens, and for pathology laboratories, this has largely involved evaluation of programmed cell death ligand-1 (PD-L1) immunohistochemical expression, in addition to a variety of other immune cell markers that are now showing correlation, such as CD8 immunostaining.

Pathological evaluation of the TME (eg, PD-L1 score, CD8 quantification) is challenging in lung cancer given that many patients have a variety of small biopsy samples obtained for diagnosis, and there is inherent heterogeneity within these tumour samples that has been shown to affect precise quantification on surgical specimens.^11,13,18 However, studies have shown that PD-L1 detection in exfoliative fluid specimens of NSCLC has a high concordance with matched histological samples, and that quantification of PD-L1 expression is feasible on cytology and small biopsy specimens with comparable results to tissue specimens.^19,20

Due to challenges with interpreting immunomarkers, and the importance of the results for prognostication and treatment selection, a few studies have started to look at the use of image analysis (IA) for evaluation of immunomarkers. However, there is a paucity of literature looking at ways to quantify immune cell markers using IA in small biopsies or tissue microarrays.^6,18 The application of IA for small biopsies in lung cancer is important, given that lung cancer is the most common cause of cancer death, and that the majority of patients with NSCLC present with metastatic disease amenable to diagnosis with small biopsies (including endobronchial ultrasound-guided transbronchial fine needle aspirations [FNAs], computed tomography- or ultrasound-guided FNAs).²¹ Thus, in this study, the utility of IA to quantify CD8 stained T cells in a series of matched small biopsies and subsequent resections of primary NSCLC is investigated to see whether CD8+ T-cell quantification on small biopsies correlates with histological findings and whether this new technology can therefore be utilised on small samples to make this more reproducible.

2 |. MATERIALS AND METHODS

2.1 |. Patients

Patients with primary NSCLC resections at our institution were evaluated over a 5-year time period (2014–2019) to find cases with pre-operative cytology cell-blocks (CBs) or core needle biopsies (CNBs) from the same anatomic location of the resected tumour. Cases were obtained using our pathology laboratory information system (CoPath, Cerner), then reviewed to find representative formalin-fixed paraffin-embedded blocks of lesional tumour, which were then requested and blank slides obtained. This study was approved by the University of Pittsburgh’s Institutional Review Board (IRB).

2.2 |. Immunohistochemistry

CD8 immunohistochemistry (clone CD8/144B, Dako M7103) was performed on tissue sections from the formalin-fixed paraffin-embedded blocks, after pre-treatment of the tissue, on the Ventana Benchmark automated slide stainer (Ventana Medical System). The antibody was diluted to a concentration of 1:40, and DAB (3,3′-diaminobenzidine) was used as the colour reagent for visualisation. All slides were counterstained with haematoxylin for visualisation of the nuclei. Appropriate positive and negative controls were utilised for all cases. The haematoxylin- and eosin-stained slides, along with the CD8-stained slides, were then digitised using an Aperio AT2 scanner (Leica).

2.3 |. Image analysis

Quantitative IA was performed with a laboratory developed, customised Aperio Nuclear v9 algorithm (Leica) to detect cytoplasmic CD8 staining, as previously described, which was previously cross validated with the fluorescent-based AQUA method.⁶ CD8 score (number of T cells with 1–3+ cytoplasmic staining/total area in mm²), which accounts for the density of CD8+ cells/mm², was determined for all samples. Any staining was considered to be positive (ie, intensity score of 1, 2, or 3). Scoring was done in two ways for each case: (a) using the entire whole slide image (WSI) and (b) analysing discrete representative fields of view (FOVs) for each case targeting tissue spots (TS) and minimising non-tissue area or problematic areas (eg, section folds, pigment; Figures 1 and 2). The TS was defined as the area of viable tumour cells with their associated stroma, as used in other studies evaluating immunomarkers.^20–24 For cell blocks, TS was the tissue fragments of areas with clusters of tumour cells. The FOV for the TS scoring was performed by a pathologist annotating the CD8-stained scanned slides, and 1–4 FOVs were selected. Then mean CD8 score for each type of specimen (CB, CNB, resection) using the different types of annotation method (WSI or TS) were calculated and used to compare all individual values to determine whether they fell above or below the mean value, given that no cutoff value is established.

FIGURE 1.

Sample annotations on cell-block (CB), core needle biopsy (CNB) and lung resection (RS) specimens. The digitised CD8-stained formalin-fixed paraffin-embedded sections of CB, CNB and RS cases (low power) were annotated to capture the whole slide image (green box) or tissue spot (yellow), fields of view with representative tumour cells contoured to limit the inclusion of non-tissue elements such as background white space.

FIGURE 2.

Representative example of each type of specimen (cell-block [CB] top panel, core needle biopsy [CNB] middle panel, resection bottom panel) using the different types of annotation methods (whole slide image [WSI] or tissue spot [TS]) for CD8 scoring in a 58-year-old man with adenocarcinoma of the lung. The heterogeneity of the CD8 staining is seen in the resection specimen. This specimen was given a high CD8 score by WSI CD8 quantification on the CB, but was correctly classified as low on the CNB and resection specimens by both WSI and TS annotation, and by TS annotation on the CB. Note: Blue is 0, yellow is 1+ staining, orange is 2+ staining, maroon is 3+ staining

3 |. RESULTS

A total of 44 patients with primary NSCLC resections at our institution with corresponding pre-operative cytology CBs (35, 79.5%) and/or CNBs (14, 31.8%) were selected. Five (11.4%) patients had all three specimen types (CB, CNB and resection), allowing for a total of 49 specimens to be reviewed with the CD8 algorithm. The patients included 33 (75.0%) women and 11 (25.0%) men with an average age of 74 years (range 43–80 years). The site of the primary lung tumours included 27 (61.4%) in the right lung and 17 (38.6%) in the left lung. Primary diagnoses on the resection specimens included adenocarcinoma (40 cases, 91.0%) and squamous cell carcinoma (four cases, 9.0%).

The mean CD8 score was determined in resections (190.35 WSI, 336.58 TS) and corresponding CBs (7.67 WSI, 77.67 TS) and CNBs (47.35 WSI, 325.67 TS). These mean values were then used to determine if the individual values for each case were above or below the average for the corresponding specimen type (CB, CNB or resection) and annotation method (WSI or TS), since there are no established cut-offs. An example of an adenocarcinoma of the lung sampled by FNA, CNB and resection is shown in Figure 2 with the CD8 stained slide analysed by IA by the WSI and TS annotation methods.

Then two-by-two tables were created to look at the concordance of the values for each annotation method, comparing CB to resection, CNB to resection, and combined CB and CNB to resection. Using these tables, the CD8 score correlated above or below average in both CB and resections for 24 (68.6%) cases, regardless of annotation method (Table 1). In the cases with WSI annotation that correlated (24 cases), six (17.1%) cases had above average CD8 scores on both CB and resection, while 18 (51.4%) had below average CD8 scores on both CB and resection. In the subset of cases that did not correlate by WSI annotation (11 cases), six (17.1%) had a higher CD8 score on the CB (overscored on CB) and five (14.3%) had a higher CD8 score on the resection (underscored on CB). Overall, the WSI CD8 score on CBs correctly identified a low CD8 score in 18 (75%) of the 24 low CD8 scores identified on resection cases (Table 2a). In the CB cases with TS annotation, there were 20 (57.1%) cases that had below average CD8 scores on both CB and resection, and four (11.4%) with above average CD8 scores on both CB and resection. In the cases that did not correlate by TS annotation (11 cases), four (11.4%) had a higher CD8 score on the CB (overscored on CB) and seven (20.0%) had a higher CD8 score on the resection (underscored on CB). Overall, the TS CD8 score on CBs correctly identified a low CD8 score in 20 (83%) of the 24 low CD8 scores identified on resection cases (Table 2b).

TABLE 1.

Clinicopathological findings of non-small cell lung cancer study cases

Clinicopathological features	Quantitative findings
Total cases	44 patients
Cell blocks	35 (79.5%)
Core needle biopsies	14 (31.8%)
Resections	44 (100%)
Age	74 years (range 43–80 years)
Sex	33 (75.0%) female
	11 (25.0%) male
Location of primary tumour	27 (61.4%) right lung
	17 (38.6%) left lung
Primary diagnosis (on resection)	40 (91.0%) adenocarcinoma
	4 (9.0%) squamous cell carcinoma
Whole slide CD8 score
Cell-blocks	7.67 average density (range 0.60–49.50)
Core needle biopsies	47.35 average density (range 7.02–287.29)
Resections	190.35 average density (range 67.58–841.43)
Tissue spot CD8 score
Cell blocks	77.67 average density (range 7.21–362.51)
Core needle biopsies	325.67 average density (range 22.58–1357.05)
Resections	336.58 average density (range 91.00–1446.00)
	Concordant (%)	Discordant (%)
Correlations based on WSI
WSI CD8 score CB vs resection	24 (68.6)	11 (31.4)
WSI CD8 score CNB vs resection	11 (78.6)	3 (21.4)
WSI CD8 score CNB/CB vs resection	35 (71.4)	14 (28.6)
Correlations based on annotated area
TS CD8 score CB vs resection	24 (68.6)	11 (31.4)
TS CD8 score CNB vs resection	7 (50.0)	7 (50.0)
TS CD8 score CNB/CB vs resection	31 (63.3)	18 (36.7)

TABLE 2.

(a) Correlation of cell-block (CB) and resection CD8 scores using whole slide image (WSI) annotation and (b) tissue spot (TS) annotation

Score	Resection WSI CD8 score (high, >average, %)	Resection WSI CD8 score (low, <average, %)	Total
(a)
CB WSI CD8 score (high, >average)	6 (17.1)	6 (17.1)	12 (34.3%)
CB WSI CD8 score (low, <average)	5 (14.3)	18 (51.4)	23 (65.7%)
Total	11 (31.4)	24 (68.6)	n = 35
Score	Resection TS CD8 score (high, >average, %)	Resection TS CD8 score (low, <average, %)	Total
(b)
CB TS CD8 score (high, >average)	4 (11.4)	4 (11.4)	8 (22.9%)
CB TS CD8 score (low, <average)	7 (20.0)	20 (57.1)	27 (77.1%)
Total	11 (31.4)	24 (68.6)	n = 35

Although there were fewer CNB cases to correlate with resections, concordance was seen in 11 (78.6%) cases using WSI annotation and seven (50.0%) using TS annotation (Table 1). In the concordant cases using WSI annotation (11 cases), two (14.3%) cases had above average CD8 scores on both CNB and resection, while nine (64.3%) had below average CD8 scores on both CNB and resection. In the discordant cases using WSI annotation (three cases), one (7.1%) had a higher CD8 score on the CNB (overscored on CNB) and two (14.3%) had a higher CD8 score on the resection using the WSI annotation (underscored on CNB). In addition, nine (90%) of the 10 cases with low WSI CD8 scores on resection were correctly identified as low on the CNB (Table 3a). In the concordant cases using TS annotation (seven cases), the majority of the cases (six cases, 42.9%) had below average CD8 scores on both CNB and resection, and 67% of the low TS CD8 scores on resection were correctly classified on CNB. In the discordant cases using TS annotation (sevem cases), three (21.4%) had a higher CD8 score on the CNB (overscored on CNB), and four (28.6%) had a higher CD8 score on the resection using the TS annotation (underscored on CNB; Table 3b).

TABLE 3.

(a) Correlation of core needle biopsy (CNB) and resection CD8 scores using whole slide image (WSI) annotation and (b) tissue spot (TS) annotation

Score	Resection WSI CD8 score (high, >average, %)	Resection WSI CD8 score (low, <average, %)	Total
(a)
CNB WSI CD8 score (high, >average)	2 (14.3)	1 (7.1)	3 (21.4%)
CNB WSI CD8 score (low, <average)	2 (14.3)	9 (64.3)	11 (78.6%)
Total	4 (28.6)	10 (71.4)	n = 14
Score	Resection TS CD8 score (high, >average, %)	Resection TS CD8 score (low, <average, %)	Total
(b)
CNB TS CD8 score (high, >average)	1 (7.1)	3 (21.4)	4 (28.6%)
CNB TS CD8 score (low, <average)	4 (28.6)	6 (42.9)	10 (71.4%)
Total	5 (35.7)	9 (64.3)	n = 14

When combining the small biopsy results overall (ie, CB or CNB), the results correlated with the resection in 71.4% cases using WSI and 63.3% cases using TS annotation (Table 1). In the cases with WSI annotation that correlated (35 cases), eight (16.3%) cases had above average CD8 scores on both CB and resection, while 27 (55.1%) had below average CD8 scores on both CB and resection. In the cases that did not correlate by WSI annotation (14 cases), there was an equal number of cases (seven cases, 14.3%) that had a higher CD8 score on small biopsy or resection. Overall, the WSI CD8 score on small biopsies correctly identified a low CD8 score in 27 (79%) of the 34 low CD8 scores identified on resection cases (Table 4a). In the cases with TS annotation that correlated (31 cases), five (10.2%) cases had above average CD8 scores on both CB and resection, while 26 (53.1%) had below average CD8 scores on both CB and resection. In the cases that did not correlate by TS annotation (18 cases), there were seven (14.3%) cases with a higher CD8 score on the small biopsy (overscored on small biopsy) and 11 (22.4%) had a higher CD8 score on the resection (underscored on small biopsy). Overall, the TS CD8 score on small biopsies correctly identified a low CD8 score in 26 (79%) of the 33 low CD8 scores identified on resection cases (Table 4b).

TABLE 4.

(a) Correlation of all small biopsy (CB or CNB) results with resection CD8 scores using whole slide image (WSI) annotation and (b) tissue spot (TS) annotation

Score	Resection WSI CD8 score (high, >average, %)	Resection WSI CD8 score (low, <average, %)	Total
(a)
CB or CNB WSI CD8 score (high, >average)	8 (16.3)	7 (14.3)	15 (30.6%)
CB or CNB WSI CD8 score (low, <average)	7 (14.3)	27 (55.1)	34 (69.4%)
Total	15 (30.6)	34 (69.4)	n = 49a
Score	Resection TS CD8 score (high, >average, %)	Resection TS CD8 score (low, <average, %)	Total
(b)
CB or CNB TS CD8 Score (high, >average)	5 (10.2)	7 (14.3)	12 (24.5%)
CB or CNB TS CD8 Score (low, <average)	11 (22.4)	26 (53.1)	37 (75.5%)
Total	16 (32.7)	33 (67.3)	n = 49a

^aThere are 49 specimens because of the 44 patients, five patients had all three specimen types.

Review of the discrepant cases revealed that some discordant cases were attributed to values near the cut-off and background detection of non-specific staining, darker staining due to tissue folds or anthracotic pigment (Figure 3).

FIGURE 3.

Examples of non-specific staining detected by the CD8 algorithm in small biopsies. (A,B) More intense staining detected in areas of tissue folds due to darker red blood cells in a core needle biopsy detected by CD8 image analysis as weak staining (A. CD8 stained core biopsy, medium power; B. Image analysis of CD8 stained core biopsy, medium power). (C,D) More intense staining detected by the CD8 image analysis in areas with anthracotic pigment, which is common in mediastinal specimens (C. CD8 stained core biopsy, medium power; D. Image analysis of CD8 stained core biopsy, medium power). Note: blue is 0, yellow is 1+ staining, orange is 2+ staining, maroon is 3+ staining

4 |. DISCUSSION

Given that patients with NSCLC frequently have small biopsies and cytology specimens obtained for initial primary diagnosis, or confirmation of metastatic disease and staging, there is interest in being able to perform theranostic markers on these small specimens.^19,20 This is particularly important now as studies have shown that an increased density of CD8+ T-lymphocytes (eg, CD8+ score) in NSCLC correlated with a significantly better prognosis.^4–13 In these studies, the strongest correlation with prognosis, independent of other variables, was the density of CD8+ T-lymphocytes within the tumour-related stroma and various cut-off points have been suggested, with no consensus as to the optimal cut-off.^7,8 This has subsequently led to a growing interest in characterising the TME in NSCLC and reliably determining the CD8 score in small samples obtained from patients with advanced disease who may not be surgical candidates. This study utilised IA to quantify CD8+ cytotoxic T cells in a series of primary NSCLCs with histological resection specimens and matched cytology and/or small biopsies in an effort to investigate the feasibility of using IA in small biopsies and the concordance between small and large lung cancer specimens.

Overall, IA was able to quantify CD8+ T cells in these small biopsies, with these results correlating with the CD8 score on resection using a dichotomised scoring scheme (high/above mean vs low/below mean) in up to 78.6% of cases. The best concordance was seen using the whole slide analysis (WSI quantification), with the highest concordance seen when comparing the resection CD8 score using WSI CD8 scores with that on CNBs (78.6%) and all small biopsies (CNBs or CBs; 71.4%). In addition, IA did a superior job correctly identifying low CD8 scores on small biopsies, with the best detection using WSI CD8 score of CNBs (90% correctly classified as low). However, IA was less reliable in the few cases with high CD8 scores. These findings illustrate the ability to use IA for CD8+ T-cell quantification, and supports that the entire slide level analysis for quantification provides the best correlation with the CD8 score on the resection specimen. Given that the tissue annotation method did not appear to provide superior concordance, the added labour time for pathologists to annotate these cases does not appear to add significant benefit for CD8 quantification in small biopsies. In addition, we know that quantification by IA can be challenging, especially near the cut-off values that are set and in cases with heterogeneity which is inherent in small biopsy sampling, and has been discussed in the literature extensively with respect to Ki67 grading in neuroendocrine tumours.^25–29

The ability to use IA for quantification of cells in the TME has been shown with PD-L1 and CD8 using different cut-offs and study designs. For the evaluation of PD-L1 expression, different cut-offs have been used for tumour expression (1%, 5% and 50%) and stromal expression (5%) in whole tissue sections and tissue microarrays using numerous FOVs representative of the tumour, illustrating the prominent heterogeneity of scoring among different assays and within tumours.¹⁸ In a study looking at the CD8+ T-lymphocyte density in head and neck squamous cell carcinomas, a CD8 density greater than 136 cells/mm² was associated with a higher median survival for head and neck squamous cell carcinomas, and the correlation was seen irrespective of the type of analysis performed by the algorithm (whole slide level analysis or analysis targeting tumour subcompartments).⁶

Another study looking at CD8 scoring used a dichotomous cut-off of 204 CD8+ cells/mm² in a series of NSCLCs in whole tumour samples, and showed that the correlation with prognosis was optimised when scoring was done with select sampling strategies (eg, random sampling of 20% of the tumour, a simulated core biopsy, or from selectively sampling the tumour centre).⁹ In our study, the resection specimens had a mean of 190.35 CD8+ cells/mm², which was used as the dichotomous cut-off (above average/high or below average/low) for the CD8 score, and the best concordance was seen with slide level analysis, as opposed to selection of FOVs with more tumour tissue and less background (TS annotation). It is interesting that the TS method did not correlate better than whole slide analysis, and this may be attributed to the fact that this annotation method may selectively choose areas with less stromal tissue representation, which could make it less ideal for the CD8 score. In addition, this method is more labour-intensive for pathologists, harder to implement in busy practices, and more subject to sampling bias or interobserver variability. The superior correlation with slide level analysis was also seen in the prior study from our institution using head and neck squamous cell carcinomas.⁶

IA technology has the potential to overcome some of the interpretation issues with PD-L1 assays and other biomarkers, including the lack of standardisation and issue of subjective interpretation impacting reproducibility. However, strict guidelines on the type of staining to quantify (eg, membranous, cytoplasmic, weak vs intense staining), the different locations to evaluate (eg, intra-tumoural, stromal, central, peripheral, leading edge), the precise cut-off values to use, and how to design the algorithm, are just some of the complexities that need to be established before this can be reliably used for routine patient care. The quantification of CD8+ T-cell density using IA has not been previously reported in the cytology literature, but many studies have used tissue microarrays, which are similar to CNB specimens in that they represent limited tissue sampling of a larger tumour.⁶ In this study, the highest correlation with the resection CD8 score was seen in CNBs alone, followed by a combination of all small biopsies (CNB or CB). It is possible that the CD8 score is more representative of the resection when performed on small tissue biopsies that have more of a stromal compartment included, over aspiration CB specimens that tend to be stromal poor from the selective tumour-rich sampling that occurs. However, in CBs with large tissue fragments, such as those acquired with new endoscopic ultrasound- and endobronchial ultrasound-guided fine needle biopsies with enhanced cutting surfaces to obtain more tissue fragments, perhaps this will allow the sampling of more stroma and improved correlation between CB-associated CD8 scores and resection-based CD8 scores. In the cytology literature evaluating exfoliative and aspiration specimens for PD-L1, there has been high concordance with matched resection specimens,^19,20 but this may reflect that PD-L1 expression in NSCLC is primarily evaluated on the tumour cells, not stromal cells. Another prior study looking at CD8+ T-cell enumeration using automated software showed that different tumour sampling in 23 primary NSCLCs yielded discordant tumour-infiltrating lymphocyte density, and suggested that this may indicate that small biopsies may not be representative of the CD8 score obtained on resections pointing to sampling issues.⁹ These factors may make stromal evaluation of the CD8+ cells more problematic in cytology cell blocks. Furthermore, given some of the artefacts seen in this study of small biopsies and cytology specimens in lung specimens, including the detection of more intense staining in tissue folds and non-specific red blood cell staining or anthracotic pigment, it is possible that close oversight of the algorithm in these cases may be needed to optimise results.

Our analysis also showed a relatively high negative predictive value, whereby a low/below average CD8 score on small biopsy correlated better with a low/below average CD8 score on resection, particularly in CNBs, than did cases on the high end of the spectrum. In fact, in the few cases with high or above average CD8 scores, only about half were detected as high on the small biopsy. However, our study is limited in cases and had fewer cases with high CD8 scores for evaluation. This has also been seen in larger population-based studies looking at CD8 scores, where low or intermediate CD8 scores outnumbered high CD8 scores, with a high CD8 score only reported in 22% out of 797 patients.^7,8 This is similar to our study where only approximately 30%–35% of patients had high CD8 scores based on resection specimens. Although IA appeared to perform better in identifying low CD8 scores, the high CD8 scores may be harder to reliably detect given that a high score may be more subject to tumour heterogeneity. The inherent sampling issues in small biopsies and cytology specimens always leave open the possibility of non-representative findings or discorrelation from resection specimens given the intra-tumoural heterogeneity with respect to the spatial distribution of CD8-positive tumour-infiltrating lymphocytes. However, given the superior performance of CNB scoring in this study and the availability of new fine needle biopsy techniques obtaining more stroma and more material, it is possible that these new biopsy techniques will provide more material for the analysis of CD8 in different stromal or tumour compartments in the future. Perhaps a cellularity or volume of tissue criteria may be necessary for cell blocks to indicate which ones may be more representative. In studies looking at PD-L1 expression in cell blocks, a minimum of 100 viable tumour cells has been used as a cellularity requirement and resulted in high concordance with PD-L1 expression in matched histology specimens.^19,20 One could theorise that when determining the CD8 score there may need to be a minimum amount of stroma or non-tumour tissue available for adequate interpretation. Overall, further investigation is needed to optimise which tissue compartments should be evaluated, to establish cut-offs or scoring parameters, and to understand the optimal way for IA to determine a reliable CD8 score that optimally correlates with prognosis.

To the best of our knowledge, this is the first study examining the use of IA for establishing a CD8+ density in cytology and small biopsy specimens from NSCLC with matched resection specimens. The findings show that IA is feasible with the best correlation using whole slide analysis for quantification on CNBs and better correlation than CBs. However, given the lower concordance with CBs, a quantitative CD8 score should be performed preferentially on resection specimens or larger tissue biopsies, opposed to CBs. The better concordance with whole slide enumeration would make implementation of IA for CD8 faster for pathologists without having to select tissue FOVs for evaluation. In the future, larger studies looking at a variety of different lung and mediastinal small biopsies and cytology samples with clinical follow-up will be of interest to determine the best quantification method for CD8 in lung cancer specimens.