Colorectal cancer (CRC) is a common global disease that is the third most common type of cancer and the second leading cause of cancer deaths worldwide (1–4). In 2018, there were 1.8 million new cases of CRC (1). Thus, CRC seriously affects and threatens global health and quality of life.

The prognosis of CRC patients depends on the tumor node metastasis (TNM) stage of the cancer and, for patients with early-stage CRC, whether they receive curative surgical interventions. CRC prognosis is also influenced by epigenetics and the TIME, both of which cause CRC heterogeneity (5). Despite the poor prognosis and high recurrence rate of CRC, potential biomarkers have helped to improve the prognosis of CRCs (6–10). It has been shown that the lipid microenvironment surrounding the tumor is associated with the occurrence, development, and metastasis of CRC (11–13). In addition, obesity is associated with increased incidence and mortality of CRC (14–16). Adipocytes (ADI) in the TIME can act directly as energy providers and metabolic regulators to promote the proliferation, invasion, and even drug resistance of CRC (14). However, the lipid microenvironment of CRC is complex, and there is considerable heterogeneity among individuals; therefore, a suitable method is needed to evaluate the lipid microenvironment of CRC patients to more accurately predict their prognoses.

In recent years, the rise of the CNN in the artificial intelligence field has allowed for extraction of pathological image features by automatic model extraction as opposed to traditional manual design (17–19). Furthermore, the CNN has promoted quantitative analysis of pathological images, as opposed to only qualitative analysis. When quantitative image data is combined with accurate clinical data and follow-up information, a systematic model of prognosis and survival can be constructed, and the postoperative status of patients can be predicted to develop a personalized treatment plan (20, 21). Kather et al. (22) used a CNN to study the auxiliary diagnoses of CRC pathological sections and extract texture features from a large number of H&E sections. The features were then input into a classifier to predict the tissue type. Using this approach, recognition accuracy could exceed 94%. Xu et al. (23) trained a deep learning CNN model on ImageNet, and it was superior to manual classification and segmentation of colon cancer (CRC) histopathological images. Deep learning can further predict the survival status of patients by counting the number of mitotic cells in the CRC image, quantifying immune cell infiltration, evaluating the degree of tumor differentiation, and characterizing the surrounding tissues (24–32).

In the present study, we segmented various tissues from H&E-stained pathological images of CRC by constructing a deep convolution network. We used this network to further analyze and predict the survival status of CRC patients according to their segmented pathological tissue image information, including TIME composition. We recommend that quantified image features are useful for predicting CRC patient prognosis and informing treatment plans.

Understanding and characterizing CRC is very important for evaluating the prognosis of patients and informing treatment decisions (47). Previous studies have shown that clinical variables, histopathological parameters, and molecular characteristics are associated with the clinical prognosis of patients (48, 49). For example, Wulczyn et al. constructed a deep learning system (DLS) score that can predict the disease-specific survival (DSS) of stage II and stage III CRC patients by analyzing H&E images. They found that CRC patients with higher DLS scores had significantly shortened DSS times (47). In addition, Zhao et al. (50) trained a CNN model using transfer learning and quantified the tumor-stroma ratio (TSR) after segmenting H&E images of CRC patients. They found that the OS time of the high-TSR group was significantly lower than that of the low-TSR group (P < 0.001, HR = 1.79). In the present study, we used a CNN model to segment CRC images from the TCGA-CRC queue and two Local-CRC queues and automatically quantified the adipose tissue scores of the H&E images. Further analysis showed that the ADI score constructed by the CNN could predict the OS of CRC patients. CRC patients in the high-ADI group had significantly lower OS times than patients in the low-ADI group. In the subgroup analysis, ADI could also be used as a prognostic marker for COAD, READ, CRC-Male, and CRC-Female patients. Across these groups, patients with lower ADI had significantly improved OS times. To our knowledge, this is the first study to establish a deep learning model based on WSI for automatic quantification of ADI. We have verified the predictive effectiveness of ADI on prognosis in several independent queues. Our method can quantify ADI from H&E-stained histological images efficiently and accurately, thus eliminating deviation caused by traditional visual evaluation and reducing the workload for pathologists.

The role of adipose tissue in the occurrence, development, invasion, and metastasis of CRC is receiving increased research attention (14). In addition to constructing a biomarker, ADI, that can predict CRC patient prognosis, we hope to clarify the molecular mechanisms by which high-ADI negatively impacts patient prognosis (Figure 7). GSEA and ssGSEA were used to compare the upregulation or downregulation of pathological pathways between high-ADI and low-ADI CRC patients. We found that anti-tumor immunity pathways, including cytokine production, chemokine signaling, TCR signaling, BCR signaling, and negative regulation of lymphocyte apoptosis, were significantly downregulated in the high-ADI group (51, 52). Kawamura et al. (53) have shown that cytotoxic T lymphocyte (CTL) responses to peptide vaccines can predict the survival time of patients with stage III colorectal cancer. Shibutani et al. (54) examined the infiltration of tumor infiltrating lymphocytes (TILs) in the primary tumors of patients with stage IV CRC, according to the method proposed by the international TILs working group. They found that patients with high TILs had a higher response rate to chemotherapy than patients with low TILs (79.3% vs. 48.1%, respectively, P < 0.025). Furthermore, patients with high TILs had higher OS times than those with low TILs (median survival time 35.5 months and 22.4 months respectively, P < 0.0221). In addition, Emile et al. (55) discussed the prognostic value of TILs in 1,220 patients with stage III CRC who received folinic acid, fluorouracil, and oxaliplatin (FOLFOX) chemotherapy. They found that the recurrence rate of the high TILs group was 14.4%, whereas the recurrence rate of the low TILs group was 21.1% (P = 0.020). Patients with high TILs had higher OS and disease-free survival (DFS) than patients with low TILs. Thus, the level of lymphocyte infiltration into the tumor microenvironment may affect the efficacy treatment regimens and can predict the prognosis of CRC patients to an extent. M1 macrophages secrete tumor necrosis factor-α (TNF-α), which can kill and inhibit the growth of tumor cells. M1 macrophages highly express major histocompatibility complex (MHC) class I and class II molecules, which present tumor-specific antigens and indirectly inhibit the growth of tumor cells (52). Therefore, differences in M1 macrophage levels could explain why patients with high-ADI CRC have significantly shortened OS times. In addition, we found that the activity of some carcinogenesis pathways, such as fibroblast growth factor signaling, angiogenesis, ROS synthesis processes, JAK/STAT signaling, and IL-1 production, was significantly upregulated in the high-ADI group compared to the low-ADI group (56–58). The activated MAPK and PI3-K pathways have been shown to regulate gene expression and protein expression, thus promoting cell growth and proliferation and reducing apoptosis (56). IL-1 and IL-6 have been shown to have multiple effects on multiple cell types, and these cytokines participate in various stages of tumor formation, invasion and metastasis. IL-1 and IL-6 can regulate the carcinogenic transcription factors NF-κB and STAT3, which play important biological functions and can accelerate tumor growth and progression (59). In addition, IL-6 can promote platelet production (60), and activated platelets can act as chemotactic agents for tumor cells, promoting the formation of metastatic lesions and increasing the level of circulating tumor cells (57, 58). High levels of VEGF can promote the growth, angiogenesis, and metastasis of tumor cells (61–63). A chronic inflammatory microenvironment can promote cell proliferation and cause uncontrolled growth, which leads the formation of tumors. Inflammation can also promote tumor progression through genomic instability, initial infiltration, and metastasis (64). In addition, studies have shown that fibroblast growth factor-2 (FGF-2) regulates tumor angiogenesis (65). Tumors often originate from sites of chronic inflammation or infection, and large numbers of phagocytes can induce oxidative stress by producing ROS and reactive nitrogen species (RNS) after infiltrating the site of inflammation, thus causing DNA damage in the host tissues. Therefore, persistent inflammation can promote the accumulation of mutations (66–68). For this reason, significant upregulation of carcinogenesis-associated pathways in high-ADI patients could explain their significantly shortened OS times.

In the present study, we found that high-ADI may worsen the prognosis of CRC, and there was no correlation between BMI and prognosis of TCGA-CRC (Figure 2D; log-rank P > 0.05). In the correlation analysis, we found that there was no significant relationship between ADI and BMI (Figure 2C, P > 0.05). At present, there is controversy about the relationship between BMI and patient prognosis (69). Many studies have shown that there is no significant relationship between BMI and prognosis (70–72); however, some studies have shown that BMI is positively correlated with mortality (4, 44). Therefore, we suspect that BMI may have a different effect than adipose tissue on the occurrence and progression of CRC. Adipose tissue is an important endocrine organ that can induce and secrete a variety of endocrine factors and fat factors, thus exerting anti-inflammatory and pro-inflammatory functions (73). Moreover, obese patients often have increased leptin, decreased adiponectin, and increased pro-inflammatory adipokines (74). Leptin activates the JAK/STAT pathway by binding its receptor (OB-R) and thereby promotes the adhesion and invasion of colorectal cancer cells (75). On the other hand, adiponectin can inhibit the proliferation of cancer cells and induce apoptosis (76).

In the present study, ADI obtained from deep learning analysis of CRC patient H&E images could be used as a potential biomarker for the clinical prognosis of a CRC patient. However, there were limitations in the present study. First, when training the model, many samples needed to be included, and the samples needed to be cut into many small squares; therefore, building the model required considerable computer space and operation time. However, when the completed model was used to interpret a patient's H&E slice, only a few seconds were needed to obtain classification results. Second, large images require large amounts of memory, and consequently, calculations are complicated and slow. For this reason, we segmented images into smaller pictures. When segmenting images, we tried our best to ensure accuracy, but some errors may still have occurred. Accuracy will therefore be improved if a pathologist performs secondary screening based on the small cut pictures. Third, the ADI constructed by the CNN model is only aimed at predicting OS, but not other prognosis metrics, such as progression-free survival (PFS), disease-free survival (DFS), and DSS. Fourth, we could not assess the association between ADI and known prognostic factors, such as tumor budding, number of lymph nodes, tumor location, microsatellite instability, TILs, mutations (such as BRAF and KRAS), or histological subtypes. Fifth, in the two independent verification sets, the predictive value of ADI for subgroup prognosis was not entirely consistent. Therefore, future studies should include a larger sample of queues for verification.

In the present study, we used a CNN model to automatically quantify ADI for H&E-stained CRC images. Furthermore, we found that ADI may predict OS in CRC patients and their subgroups (COAD, READ, CRC-Female and CRC-Male). The present study shows that automatic histopathological image analysis can be achieved through a deep learning model. The quantified image features could assist with predicting patient prognosis and guide clinical decision-making.

The original contributions presented in the study are included in the article/Supplementary Material, further inquiries can be directed to the corresponding authors.

The studies involving human participants were reviewed and approved by Zhujiang Hospital of Southern Medical University and First People's Hospital of Chenzhou. The patients/participants provided their written informed consent to participate in this study. Written informed consent was obtained from the individual(s) for the publication of any potentially identifiable images or data included in this article.

PL and JZ: conceptualization. AL, CQ, and ML: formal analysis and visualization. AL, CQ, ML, RG, QC, ZL, XW, QL, JZ, and PL: writing—original draft. AL, CQ, ML, RG, EI, NM, QC, ZL, XW, QL, JZ, and PL: writing—review and editing. All authors read and approved the final manuscript.

This work was supported by the Natural Science Foundation of Guangdong Province (Grant Nos. 2018A030313846 and 2021A1515012593), the Science and Technology Planning Project of Guangdong Province (Grant No. 2019A030317020), and the National Natural Science Foundation of China (Grant Nos. 81802257, 81871859, 81772457, 82172750, and 82172811).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.