Locally advanced rectal cancer (LARC) is usually treated with neoadjuvant chemoradiotherapy (nCRT), followed by total mesorectal resection. Response to nCRT largely determines the prognosis and survival of patients with LARC (1). Patients may achieve pathological complete response (pCR), which is defined as the complete absence of tumor cells in the resected specimen, while some patients may only have partial response or no response. The reported pCR rate of LACR to nCRT is relatively low, ranging from 10% to 30% (1–4). Patients who achieve a pCR may adopt a conservative “watch and wait” strategy without surgery (5). Therefore, the extent of response to nCRT affects clinical decision-making and determines whether patients should be directed to aggressive surgical treatment. Currently, the gold standard for pCR relies on pathological confirmation of a surgical specimen. There is a need for developing non-invasive methods to reliably predict the response to nCRT and to avoid surgery for patients with pCR.

Predictors of pCR include clinical demographics, tumor morphology, blood cell counts, serum oncological indicators, protein expression, gene profiles, conventional radiological imaging features, the time interval between nCRT and surgical resection. However, these predictors have produced inconsistent results with some indicating their usefulness in predicting pCR and some indicating otherwise (2, 6–8). In addition to CT imaging, current efforts and studies with other imaging modalities have been performed to assist in prediction of pathological responses with promising results. PET/CT has the advantage of detecting the metabolically active rectal cancer and metastases for staging; and semiquantitative parameters derived from sequential PET/CT imaging may be used to predict response (9, 10). Endorectal ultrasound (EUS) affords direct visualization and access for biopsy of rectal cancer and its adjacent lymphadenopathy, which has also been used to predict response to nCRT (11). MRI may provide exquisite anatomical details on tumor morphology and its association with adjacent structure (12, 13). However, these additional imaging modalities are more time-consuming and costly as compared to CT imaging. In addition, some patients may not be able to tolerate the endorectal probe for EUS or the endorectal coil for rectal MRI. As CT remains to be the most commonly used imaging modality in clinical practice, it is prudent to assess its potential for predicting pCR in patients with LARC.

Radiomics detects tumor image features through computational analysis, which may reflect biological properties of tumors. It has been used to predict treatment response in patients with gastrointestinal tumors including rectal cancer (14). Prior studies have shown robust performance of predictive models for pathological response, achieving the area under the curve of 0.98 in models combining radiomic signature and clinical parameters such as tumor length (15). Radiomic features reflecting tumor heterogeneity such as entropy, skewness and kurtosis have been indicated as most relevant for predicting response to treatment (14). However, few radiomic studies have focused on these relevant radiomic features to determine tumor heterogeneity for predicting pCR in patients with LARC (2, 8, 16). In addition, prior radiomics studies of rectal cancer had small sample sizes without external validation, predisposing to overfitting and issues of generalizability (17). More work is needed to develop non-invasive radiomic approaches for cancer diagnosis and treatment.

Pretreatment contrast-enhanced CT for radiotherapy treatment planning is routinely acquired for patients with LARC prior to nCRT, which has provided a platform for predictive modeling of clinical response through imaging analysis. However, there have been conflicting results using radiomics to predict response to treatment, with some studies indicating its usefulness and some studies indicating otherwise (2, 8, 18–21). In the present study, we analyzed pretreatment contrast-enhanced CT images acquired for radiotherapy treatment planning, extracted radiomic features, and incorporated clinicopathological risk factors to build models for predicting pCR in patients with LARC undergoing nCRT. We hypothesized that the combined prediction model built with both radiomics and clinicopathological parameters could be useful for differentiating patients with pCR from patients without pCR.

In this study, we performed radiomic analysis on pretreatment radiotherapy planning CT images performed for routine clinical care and developed a radiomic nomogram to predict the response to nCRT in patients with LARC. Our predictive model, built with a combination of radiomics and clinicopathological risk factors, had reasonable performance in differentiating patients with and without pCR. The clinical usefulness of the model was also confirmed by the decision curve analysis. Our study results support the notion of using routine radiation treatment planning CT imaging to predict treatment response in clinical practice.

There is limited literature on the use of CT-based radiomics to predict the response to nCRT in patients with LARC. A prior study using a deep neural network method based on contrast-enhanced CT images achieved moderate accuracy of 80% for predicting pCR in patients with LARC receiving nCRT (16). In addition, neural network methods are difficult to understand intuitively and are considered “black box” methods for their lack of interpretability. On the other hand, our model was built with radiomic features derived directly from the tumor images, making it more relatable to clinical practice. Moreover, the prior study using deep neural network methods had a sample size of 95 subjects (16) and our model had the advantage of training and validating on a larger cohort of 216 subjects. Larger sample sizes are critical for model building because of the need to avoid overfitting of the prediction models.

Our study used contrast-enhanced CT images for radiomic analysis, which should be superior to using non-contrast enhanced images because contrast enhancement should reveal more details on the heterogenous internal architecture of malignant tumors. Two prior studies used a non-contrast-enhanced CT radiomic model to predict pCR in rectal cancer and their results were conflicting (2, 8). The model reported by Yuan et al. (2) had 83.9% accuracy and promising predictive power, while the model reported by Hamerla et al. (8) showed no predictive power for treatment response. Our results support the use of radiomics based on contrast-enhanced CT images in predicting treatment response of LARC.

CT has been the most commonly used imaging modality for diagnosis and treatment planning of various diseases. For instance, a recent study presented the feasibility of using CT images imported into a radiation treatment planning system for volumetric assessment of COVID-19 pneumonia lesions and for localizing the pulmonary lesions as a target for 3D conformal radiation therapy (31). This innovative approach is encouraging and may have potential for treatment planning of rectal cancer since nCRT also requires identification of the tumor location and volumetric assessment for radiation therapy. More work needs to be done to assess this innovative method in diagnosis and treatment of rectal cancer.

Our model combining the Rad-score with clinical features achieved robust performance. These results support further investigation of the Rad-score as a biomarker to predict the response of LARC to nCRT. Radiomic features might reflect the biological nature of tumor cells and have been found to be associated with tumor prognosis (32, 33). For instance, higher heterogeneity may indicate a more aggressive tumor with a worse prognosis while lower heterogeneity might indicate more angiogenesis with a better prognosis (33). In our study, two entropy-related features (WavEnLH_s_3, WavEnLH_s_4, see Supplemental file), indicating higher heterogeneity, were among the most relevant radiomic features filtered by the LASSO method. Both had negative coefficients, implying higher values of these features being associated with reduced likelihood of pCR. Our results were consistent with a model wherein higher entropy reflecting higher heterogeneity and worse prognosis.

The mechanism underlying the relationship between the location of rectal cancer and response to treatment is not clear yet. Our study included tumor location aiming to assess this relationship further. Prior studies have suggested that distal location of rectal cancer tended to have better prognosis, either in terms of pCR rate or in distant metastasis-free survival (7, 34–37). One study found that tumors located higher vs. lower in the rectum had pCR rates of 23.2% vs. 7.3%, respectively (34). Another study reported a similar finding, with pCR rates of 20% and 3%, for tumors higher and lower in the rectum, respectively (37). However, one study found a higher pCR rate in lower tumor location in the rectum (38). Our results were consistent with the majority of published literature showing a higher pCR rate in the tumors located higher in the rectum. We speculate that the differential blood supply, the anatomic structure, and the density of the surrounding tissue might affect the chemotherapeutic drug concentration and radiation dose in the target rectal area, resulting in different responses to treatment.

Serum inflammatory cytological biomarkers as well as tumor markers were widely studied in rectal cancer (39–41). Our model included an inflammatory cytological biomarker (i.e., LMR) and a tumor marker (i.e., CEA). In a study by Li et al. (40), LMR was found to be the most relevant factor for pCR in rectal cancer, with an AUC of 0.913. Diakos and colleagues (42) proposed that systemic inflammation may affect the tumor response through the complex interplay between local immune responses and systemic inflammation. Our results support the role of LMR in assessing treatment response of LARC. CEA is a known marker of tumor response and prognosis in gastrointestinal cancer. CEA is abnormally expressed in various malignant tumors, and higher expression has been linked to worse prognosis in colorectal cancer (43). Thus, our findings that CEA was negatively correlated to pCR were consistent with the literature. However, there were some differences in CEA cut-off determinations among the studies. For example, the cut-off value to indicate CEA-positive status was set at 5 ng/mL in some studies (44, 45), and at 2 ng/mL in other studies (46). The pCR has been linked to either the post-treatment CEA level (46, 47), the pretreatment CEA level (45, 48), or both the pre- and post-treatment CEA levels (44). Nonetheless, the potential usefulness of CEA as a prognostic biomarker for rectal cancer has been recognized and our results add to the evidence that the pretreatment CEA level may help to predict treatment response.

The prognostic factors for non-metastatic rectal cancer patients have been noted to include age, nutritional condition, tumor stage, tumor differentiation, and surgery, which may independently affect overall survival (49). Our best performing model for predicting response to nCRT incorporated both radiomics and clinical variables, such as radiomic features, distance to anal verge, lymphocyte-to-monocyte ratio (LMR), and carcinoembryonic antigen (CEA). It should not be surprising for this model to perform well as radiomic features may reflect heterogeneity of tumor cells and have been found to be associated with tumor prognosis. The location of tumor (distance to anal verge), and serological markers such as LMR and CEA are associated with biological behavior of the tumor, which may affect tumor stage and prognosis. Other prognostic factors are related to the patients’ general well-being such as frailty and nutritional status. For instance, the loss of skeletal muscle mass during nCRT has been shown to be related to lower survival for patients with rectal cancer (50). Higher BMI tends to have better survival for patients with colorectal cancer (51). More research should be performed focusing on the potentially modifiable factors such as muscle mass and body weight to enhance response to nCRT.

There were several limitations to this study. First, this was a retrospective study with uncontrolled confounding variables, such as clinical staging, nodal status and time interval between treatment initiation to surgery, making case selection bias possible and affecting the performance of our predictive model. Second, this retrospective study included a modest sample size of 216 patients from a single institution over a long enrollment period, with only 44 patients achieving pCR. The modest sample size in our study might have increased the risk of overfitting the models, and a single institution data may limit the model’s generalizability. Therefore, external validation is mandatory and should be performed in future studies. Third, differences in imaging protocols may hinder the general applicability of our results. For instance, we performed contrast-enhanced CT for radiotherapy treatment planning while patients were in supine position and other centers may use non-contrast enhanced CT imaging while their patients may assume either supine or prone position. In addition, contrast-enhanced CT may differ from facility to facility depending on the experience and expertise of the clinical practice. Furthermore, a combination of limited sample size and lack of radiomics on the CT scans acquired post-nCRT but before surgery may have missed some useful features for model building to predict response to treatment. Therefore, future studies should adopt a standardized imaging protocol among all participating institutions and should have sufficient statistical power to tease out the effect of imaging protocol on model performance. Lastly, several important variables such as gene profiles, and membrane protein biomarkers were not assessed in our study, which could have helped in assessing the potential for personalized treatment and their effects on the model performance. Future prospective multicenter studies with a large sample size are needed to validate our results and assess the clinical applications of our predictive model.

This study demonstrates the usefulness of pretreatment CT radiomics for predicting the response to nCRT in patients with LARC. Our data supported the notion of using non-invasive imaging-focused approaches to assess treatment response and to guide personalized treatment for patients with rectal cancer.

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 the Ethic Committee and Institutional Review Board in Xiangya Hospital of Central South University, P. R. China (IRB No.201610070). Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements.

YM and XY: conceptualization, methodology development, and original draft preparation. QP, HY, and HLi: data analysis. CC, YF, HLiu, GG, and HY: resources collection. PP and HLin: scanning guidance and data collection. BX and HZ: draft revising and fund providing. BC: methodology supervision and draft revising. All authors contributed to the article and approved the submitted version.

This study was funded in part by the National Natural Science Foundation of China (81701847), the Natural Science Foundation of Hunan Province, P.R. China (2017JJ3497, 2018JJ2641), the Xiangya-Peking University Wei Ming Clinical and Rehabilitation Research Fund (No. xywm2015I35), and China Post-doctoral Science Foundation (2018M632997).

Author PP and HL were employed by the company GE Healthcare.

The remaining 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.