Background
The immune system was an essential component of the tumor microenvironment (TME) in colorectal cancer (CRC) [
1]. It played a central role in tumorigenesis and progression, affecting the treatment and prognosis of CRC [
2]. The need to go beyond the tumor-node-metastasis (TNM) staging system has been addressed by detecting the tumor immune microenvironment (TIME), which was affected by the type, density, and location of tumor-infiltrating lymphocytes (TILs) [
3‐
5]. The Immunoscore®, calculated based on the density of CD3
+ and CD8
+ T cells in the tumor core and invasive margin, has been shown to provide superior prognosis to TNM stage in CRC [
4,
6]. However, most of the current immune scores needed to be stained with immunohistochemistry (IHC) or other special staining (such as multiplexed immunofluorescence), which is not commonly used in clinical practice, hindering their widespread clinical application.
In most cases, high-quality hematoxylin and eosin (HE)-stained slide was sufficient to confirm the diagnosis [
7]. Immune-related features within HE-stained whole-slide images (WSIs), such as TILs and Crohn-like lymphoid reaction (CLR), can be quantified using deep learning. HE-stained slides also have immunological information, with a larger sample size and lower cost. Deep learning has recently entered the field of computational pathology and shows excellent promise for task automation [
8]. Deep learning with digital pathology has been successfully applied to breast, prostate, lung, and CRC [
9‐
15]. The tissue-level components of CRC, such as tumor-stroma ratio (TSR) and CLR [
16,
17], can be quantified using deep learning. Therefore, the artificial intelligence (AI)-based method has the potential to quantify the tissue composition and immune status using HE-stained WSIs.
The spatial distribution of TILs was important for CRC prognosis. TILs can interact with the tumor through direct contact or cytokine signaling to produce tumor-killing immune cells for protection of the organism, mainly in the interstitial region [
18]. The immune cells in the stroma are produced by the surrounding lymphoid follicles or migrate from the blood to the tumor area. As the stroma increases, the number of immune cells in the stroma decreases and the anti-tumor effect decreases [
16]. In addition, the distribution of immune cells in the stroma impacted the prognosis of CRC [
19]. We hypothesized, therefore, that a comprehensive consideration of the stroma proportion and the immune cell infiltration in stroma would further refine the prognostic stratification of CRC patients.
The aim of this study was two-fold. First, we proposed a deep learning-based immune index, the Deep-immune score, quantifying immune infiltration interaction with the stroma in HE-stained WSIs. A further investigation of its prognostic value is performed in CRC patients from three centers.
Discussion
The AI-based method could quantify the tissue composition in TME with HE-stained WSIs of CRC. To understand the basis of tumor heterogeneous clinical behavior, many scholars have focused on TME [
22,
23]. Studies have shown that TME characterization provides additional insight into the prognosis of patients with solid tumors [
24,
25]. The stroma of TME was the focus of studying the prognosis of CRC. Our previous analysis and studies by other scholars have shown that in CRC patients, abundant stroma in tumor tissue was associated with poor prognosis [
16,
26]. Results of our present work also suggested that stroma proportion quantified in HE-stained WSIs can help stratify risk of CRC patients. Patients who have Deep-TSR-high scores have a much lower 5-year survival rate than those who have low scores (67.3% vs. 78.9%). Remodeling of stroma can serve as a physical barrier to prevent tumor cells from coming into contact with immune cells [
27‐
29]. The mean stroma-CD3 density in the Deep-TSR-low group was 1350 cells/mm
2, higher than that in the Deep-TSR-high group (1011 cells/mm
2). Additionally, the stroma contains special connective tissues such as fibroblasts, mesenchymal stromal cells, osteoblasts and chondrocytes, along with extracellular matrix. The endothelial cells within it provide nourishment for tumor growth, constitute a pathway for metastatic spread through angiogenesis, and lead to resistance to chemotherapy and radiation therapy [
23,
30,
31]. Therefore, the more stromal components, the lower the OS of patients.
Furthermore, the stroma was incredibly intricate. For example, desmoplastic reaction was classified as immature, intermediate, or mature according to the different connective tissue-promoting reactions in the stroma [
32]. Moreover, according to cancer-immune phenotypes, anticancer immunity in humans can be categorized into three main types: the immune-desert, the immune–excluded, and the inflamed phenotypes [
33]. Studies have shown that the content and density of TILs in the stroma were also attached to OS [
15,
34]. A patch-level segmentation was performed in our work without dissecting each lymphocyte with precision, which cannot accurately quantify the density and spatial location of lymphocytes. However, we noted that the class of lymphocytes, one of the tissue categories in our model, were structured made up of clusters of TILs. Therefore, we tried to take the result of stroma segmentation as ROI and defined the mean predictive probability of the ROI for this category of LYM as the Deep-TIL score. Besides, we found that the mean stroma-CD3 density of the high-score group was 1.5 times higher than the low group. The automatic quantification of the Deep-TIL score could reflect the immune cells infiltration in the stroma region. Survival analysis showed that the Deep-TIL score could stratify the prognosis of CRC. The higher the Deep-TIL score, the longer the survival time. The 5-year survival rate was recorded for 70.2% of patients with a low score, 75.7% of patients with a middle score, and 85.4% of patients with a high score. This scoring method was kind to use, and this method only needed the label of the patch, which was less computationally intensive. More worth mentioning was that it also took into account the spatial distribution of TILs in the stroma.
Tumor growth pattern, aggressiveness, metastasis, and patient prognosis are the result of a combination of multiple factors. These include the interaction between components of TME from the cellular level to the tissue level [
35,
36]. Based on this, after completing the above two scores, the Deep-TSR score, and Deep-TIL score, we raised the conjecture whether the combination of the two scores could reflect more prognostic information. Patients with the highest score had the most favorable OS (unadjusted HR for score 4 vs. score 1: 0.27). Similar results were also found in the validation cohort, which revealed that our score was robust. Furthermore, we found that the full model, including Deep-immune score and clinicopathological factors, had a higher prognostic value than a clinicopathological model (iAUC, 0.726 vs. 0.713). Combined with clinicopathological factors, the prognosis of patients with CRC could be evaluated in a more comprehensive and integrated manner. We also observed that with the increase in the Deep-immune score, the stroma-CD3 density also increased. CD3
+ T cells are membrane markers of mature T lymphocytes that can be used to quantify the total number of T lymphocytes [
19]. When CD3
+ cells was increased, it represented a higher tumor lymphatic infiltration and a higher amount of tumor-killing immune cells, which has a protective effect on the organism [
37,
38]. This result supports the idea that our proposed Deep-immune score may be sufficient to predict prognosis of CRC. In addition, the Deep-immune score, which was fully automated and with HE-stained as a routine staining method and IHC-stained as a special staining, has certain economic benefits. There have been many studies suggesting that TSR and TILs could predict prognosis of other solid tumors. Take breast cancer as an example. Studies on TSR found a significant association between high tumor stroma content and poor prognosis [
39,
40]. The results of related studies on TILs showed that increased TILs concentrations were associated with increased frequency of adjuvant chemotherapy responses in all breast cancer subtypes, and they were also associated with longer survival in patients with triple-negative breast cancer and HER2-positive breast cancer [
41,
42]. The results were similar to those of our study in CRC. This suggested that if we had segmented breast cancer tissues and defined the corresponding types of tissues, using our method to calculate tumor-stroma ratio, Deep-TIL score, and Deep-immune score, could also predict prognosis for breast cancer patients. Therefore, our method has the potential ability to be applied to other solid tumors.
In stage II patients, neither Deep-TSR nor Deep-TIL score can distinguish between high-risk and low-risk CRC individuals (all P > 0.05). However, the composite score can stratify patients' prognostic risk (P = 0.018). This indicated that Deep-immune score had the potential to guide clinical risk stratification of patients with stage II CRC, which in turn could influence clinical decision-making.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.