Background
Lung cancer is the leading cause of cancer-related deaths worldwide [
1], and non-small-cell lung cancer (NSCLC) is the most common subtype of lung cancer, which comprises 85% of total diagnoses [
2]. Surgery is the recommended treatment for resectable NSCLC [
3], whereas 30–55% of patients develop recurrence and die despite the resection [
4]. Precise risk assessment is crucial for developing individualized treatment strategies. The American Joint Committee on Cancer (AJCC) tumour-node-metastasis (TNM) staging system [
5] is widely used for risk stratification, but patients prognosis varies within each stage due to biological heterogeneity [
2]. Prediction models combining the TNM stage and clinicopathologic prognostic factors (e.g. histologic type, and treatment-related factors) have improved the clinical validity of risk stratification, but the predictive performance is unsatisfactory [
6‐
8]. A novel prognostic biomarker that characterizes the biological behaviour may improve the validity of risk stratification in NSCLC.
Recent tumour biological studies have implied that the interaction between the tumours and microenvironment is associated with tumour development, invasion, metastasis, and outcome [
9,
10]. Tumour-infiltrating lymphocytes (TILs) within the microenvironment has been reported to be the prognostic factor of resected NSCLC [
11], among which T-cells (CD3
+), especially cytotoxic T-cells (CD8
+), play important roles in antitumour immunity [
12,
13]. In recent years, many studies have attempted to characterize the in situ immune infiltration based on the density of various T-cells subsets (e.g. CD8
+, CD3
+, CD4
+, FOX-P3
+, CD45RO
+, etc.) [
10,
14]. However, a generally accepted immune scoring system for NSCLC is still unavailable since there is no consensus regarding the selection of T-cells subsets and the cell quantification approaches [
13,
14].
Immunohistochemistry (IHC) on tissue sections is a simple and reliable method to identify CD3
+ and CD8
+ T-cells. The conventional method for quantifying positive cells is through manual counting performed by pathologists, which is time-consuming with poor reproducibility. There have been prior attempts at automated histopathological analysis based on NSCLC tissue microarrays (TMA), such as evaluating the density and spatial arrangement of TILs [
15], and quantifying the different subsets of T-cells [
16,
17]. However, the selection bias of TMAs may lead to high inter-observer variability [
18]. In comparison, computer-aided analyses based on digitalized whole-slide images (WSIs) evaluate the whole tissue sections without subjective selection of regions for analysis, which improve reproducibility across users, and the spatial heterogeneity within the tumour microenvironment could be better characterized [
19]. Automated workflows for evaluating the immune infiltration on IHC-stained WSIs are expected to improve the validity and reliability of NSCLC risk stratification [
20,
21], but such an algorithm remains to be developed.
This study aimed to achieve the following objectives using 2 retrospective cohorts of patients with resected NSCLC. Firstly, we developed an automated workflow for evaluating the density of CD3+ and CD8+ cells in the tumour regions on IHC-stained WSIs. Secondly, we proposed an immune scoring system based on the automated assessed cell density. We hypothesised that the integration of this immune scoring system into clinicopathological risk factors would improve the prognostic stratification in resected NSCLC.
Discussion
In this study, we developed an automated workflow for evaluating the density of CD3+ and CD8+ cells in the tumour regions on IHC-stained WSIs of NSCLC, and further proposed an immune scoring system “I-score” based on the automated assessed cell density. The generalizability of this automated workflow and novel scoring system was validated in an external independent cohort. To the best of our knowledge, this is the first study that utilized automated whole-slide images assessment of tumour-infiltrating CD3+ and CD8+ T-cells for the prognostic stratification of resected NSCLC.
The past 10 years have seen remarkable progress in medical artificial intelligence, promoting the development of digital pathology. Digital pathology implies not only the digitization of tissue sections, but also the automated assessment workflow with high validity and reliability. The application of WSIs has expanded the scope of histopathological analyses to a whole-slide level, which places higher demands on automated algorithms. Some earlier pioneering WSI-based studies predicted the prognosis of NSCLC based on automated derived image features (e.g. Haralick texture features, radial distribution of pixel intensity, etc.) [
27], or predicted the classification and mutation status using end-to-end deep learning models in a data-driven manner [
28], which had limitations in biological interpretability.
Analysing the tumour microenvironment at the tissue and cellular levels depends on precise segmentation and identification methods, but the high histologic heterogeneity in NSCLC presents a challenge to algorithm development [
29]. This study optimized the automated positive cells assessment algorithm in the following two aspects. In the tissue segmentation process, we used a semi-automated interactive approach combining the automated algorithm and the experience of pathologists. The tumour region was determined by precisely removing adjacent normal tissues, blanks, and backgrounds to reduce the errors in estimating the tumour area. The tumour-adjacent atelectasis (belongs to normal tissue) was easily confused with tumour-associated stroma (belongs to tumour region) in this thresholding segmentation framework, so the experience of the pathologist was dispensable for identifying these tissues. The blank area (residual alveolar cavity) was a unique structure for lung cancer tissue sections, and its size varied with histologic subtypes [
30]. In previous studies, the density of positive cells was defined as the counting of positive cells per unit area (mm
2) [
31], and the area could be the high power field [
32] or the tissue surface area [
14]. Some other studies defined the density as the percentage of positive cells among total nucleated cells [
14,
33]. Our study calculated the density of positive cells using tissue surface area as the denominator, and the evaluation would be robust across histologic subtypes. As a result, the I-score based on the density of CD3
+ and CD8
+ T-cells showed good stratification performance in the adenocarcinoma and squamous cell carcinoma subgroup (Additional file
1: Figure S4d, e). In the cell segmentation process, dust macules (similar to, but slightly darker than positive cells) were filtered out to avoid being mistakenly identified as positive cells. As a result, there was a good agreement between manual counting and automated counting using our algorithm (ICC, 0.91).
Although for colon cancer, there has been a well-developed workflow for WSI assessment of Immunoscore [
18], a generally accepted immune scoring system for NSCLC prognostic stratification is still unavailable. Selecting which types of TILs and which regions/compartments of TILs for scoring has always been controversial. We referred to the findings of previous Immunoscore-related studies on NSCLC [
13,
14], and selected CD3
+ (pan T-cells) and CD8
+ (cytotoxic T-cells), two robust prognosis-associated markers in various solid cancers including NSCLC [
10,
31], for quantitative assessments. Concerning the regions for cell quantification, some studies (especially TMA-based studies) quantified the positive cells in the central tumour and the invasive margins respectively [
33,
34]. Instead, we constructed the immune scoring system based on the positive cell density in the entire tumour regions (tumour nests) on WSIs, as in some previous studies [
14,
35]. Therefore, the characteristics of immune infiltrations in the central tumour and the invasive margins (if it existed on a WSI) had been taken into account, and the selection bias could be reduced.
The I-score (two-category) that integrated the CD3-score and the CD8-score was associated with DFS after adjusting for TNM stage and other clinicopathologic factors. This finding was verified in an external validation cohort with significant differences in baseline characteristics compared with discovery cohort, suggesting that the I-score obtained by the automated workflow was an independent and robust prognostic factor of DFS in resected NSCLC. Furthermore, the prognostic value of the I-score was confirmed in the vast majority of subgroups (Additional file
1: Figure S4). The predictive accuracy (iAUC and C-index, C-index: 0.588 vs. 0.58 for validation cohort) of the I-score was similar to that of the Immunoscore of colon cancer [
36]. By integrating the I-score (two-category) into the TNM stage model and clinicopathologic model, respectively, the models with I-score showed better discrimination and calibration than those without I-score in both cohorts (Fig.
5b), which suggested that the I-score based on the automated assessed cell density would improve the prognostic risk stratification in resected NSCLC. Also, the full model yielded better discrimination compared with the reported prediction models that involved only clinicopathologic prognostic factors [
6‐
8] (C-index, 0.695 vs. 0.67, 0.664, 0.66 for validation cohort).
As for the I-score distribution across TNM stages, an interesting trend was found that a low I-score was significantly associated with the advanced TNM stage. We speculated that this might be attributable to the evolution of immune escape. A similar finding was reported in a recent genomic study on the spectrum of immune infiltration from preneoplasia to invasive lung adenocarcinomas [
37]. Still, the underlying mechanism of these findings warrants further investigation.
This study has limitations inherent to most retrospective studies. The clinical validity of this automated workflow and immune scoring system needs to be further validated in larger prospective cohorts. Besides, the quality control of WSIs was performed manually, and some parameters for tumour region segmentation were fine-tuned, if required, according to the pathologists’ proofreading. Based on the findings in this study, we are currently developing a deep-learning framework to perform NSCLC tissue segmentation, which would enable automated segmentation and identification of tumour regions and tumour-associated stroma. The density of CD8
+ cells in the stroma compartment was reported to be an independent prognostic factor in resected NSCLC [
10,
33,
38], and a precise segmentation algorithm would be an essential prerequisite for evaluating immune infiltration in the stroma compartment.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.