Introduction

According to 2021 statistics, colorectal cancer (CRC) is the third most common cancer and is the third leading cause of death among all types of cancer, accounting for about 9% of total cancer mortality in men and 8% in women [1]. Microsatellite instability (MSI) status offers a predictive value in response to adjuvant chemotherapy and overall prognosis in CRC cancer [2, 3]. MSI high (MSI-H) colorectal cancer has a better prognosis, in part because it is more likely to benefit from immunotherapy. The National Comprehensive Cancer Network (NCCN) guidelines recommend MSI testing in all CRC patients [4].

MSI status is usually determined by pathological examination of tumor samples. However, the heterogeneity of MSI status within the tumor tissue or between different sites of the tumor, as well as the heterogeneity during different stages of the tumor treatment process, leads to limitations in histological examination methods. Furthermore, histological evaluation of MSI status is rather complex and may not be achieved in some centers without delay, including the transfer of samples to a specialized center. Around 2% of rectal tumors are MSI-H [5]. It might be helpful to develop a relatively non-invasive and convenient surrogate biomarker to predict MSI status.

Radiomics studies in rectal cancer, using magnetic resonance imaging (MRI) mainly reported about the prediction of treatment efficacy [6, 7], tumor stage [8], tumor metastasis [9, 10], and patient survival [11]. There are few MRI-based radiomics reports predicting MSI status in rectal cancer [12, 13]. These studies were conducted in single centers and lacked external validation. The superiority of the multi-sequence MRI in the prediction of MSI status in CRC patients remains to be confirmed. Therefore, the aim of our study was to develop models based on a single sequence and multiple MRI sequences and to evaluate the generalizability using external validation.

Materials and methods

Patients

From May 2017 to May 2020, a total of 100 rectal cancer patients from hospital I were enrolled as the training dataset and 99 patients from hospital II and hospital III enrolled as the external verification dataset in our study. The inclusion criteria were as follows: (1) histopathological diagnosis of rectal adenocarcinoma; (2) absence of any type of anti-cancer treatments before MRI examination; (3) time interval of less than one week between MRI examination and surgery; (4) complete MRI examination sequences, including T1WI, T2WI, DWI, and CE-T1WI, in particular, DWI sequences with same or similar b values; (5) available MSI status as confirmed by IHC staining. The exclusion criteria were as follows: (1) incomplete clinical data; (2) insufficient MRI image quality or incomplete sequences; (3) lesions whose identification was difficult on MRI images.

Clinical data, including age, gender, location of the tumor (upper, middle, and low), carcinoembryonic antigen (CEA) (normal or abnormal), and carbohydrate antigen-199 (CA199) status (normal or abnormal), were recorded.

MSI status assessment

The MSI status was confirmed by the expression levels of four MMR proteins, namely MLH1, MSH2, MSH6, and PMS2, which were identified by IHC staining. Specimens were fixed by formaldehyde, and they were paraffin-embedded and routinely stained with hematoxylin-eosin (HE). The normal intestinal mucosa and inflammatory cells present in each section were used as internal controls to interpret the IHC staining results. The results were read and confirmed by two experienced pathologists specializing in CRC diagnosis for 10 years. Based on the expression of MMR proteins, the patients were divided into the following two groups: the MS-low/stable (MS-L/S) group (positive staining of four MMR proteins) and the MSI-H group (any one of the MMR proteins was negative).

MRI protocol, preprocessing, and segmentation

MRI protocol

All MRI examinations on the enrolled patients were performed using a 1.5-T scanner (Optima or Signa, GE Medical Systems). The imaging sequences included fast spin echo T1-weighted sequence, T2-weighted sequence, DWI, and contrast enhancement. Imaging acquisitions of the sequences used in the research are described in Appendix E1.

Tumor segmentation

The Dr. Wise Multimodal Research Platform (https://keyan.deepwise.com,V1.6.2) (Beijing Deepwise & League of PHD Technology Co., Ltd.) was used for radiomics analysis, which included image annotation, feature extraction and selection, and model construction. Among them, the feature extraction module was implemented with the open-source PyRadiomics library.

(https://github.com/Radiomics/pyradiomics). The workflow of the radiomics model construction and validation is shown in Fig. 1.

Fig. 1
figure 1

Schematic shows the workflow for this study

Full size image

All images were processed in the original DICOM format and transferred to a post-processing workstation, where two experienced abdominal radiologists manually annotated the lesion in the original image to obtain a 3D volumetric map of the ROI for feature extraction and quantification. Tumors were delineated along the border of the lesion in the images of each sequence (Fig. 2), covering the surrounding burrs while excluding the bowel contents and adjacent mesenteric fat tissue, as well as the non-invaded rectal wall. Both radiologists were blinded to the MSI status. A discrepancy was then resolved by discussion and consensus. Additionally, a total of 30 patients were randomly selected to evaluate the intra-reader and inter-reader agreement. Reader 1 performed the segmentation of the 30 tumors twice with a 1-month interval. Reader 2 performed the segmentation of the 30 tumor features once. Features with an inter-class correlation coefficient > 0.75 and an intra-reader correlation coefficient > 0.75 were considered to be satisfactory agreement and then were selected for further analysis.

Fig. 2
figure 2

Axial T1WI, T2WI, DWI, and CE-T1WI MR images of rectal cancer with MS-L/S type (a–d) and MSI-H type (e–h) and the corresponding manually region of interest (ROI)

Full size image

Feature extraction

For images with different resolutions, resampling was performed for image pre-processing using B-spline interpolation and all images were resampled to the same resolution [1]. The image gray-scale values were normalized. The normalization procedure was based on the following mathematical formula:

$$ f(x)=\frac{s\left[x-\mu (x)\right]}{\sigma (x)}, $$

where f(x) is the normalized image density, x is the original image density, μ(x) is the mean image density value, σ(x) is the SD of image density, and s is the scaling factor, which was 100 in this study. The absolute gray-level discretization with fixed bin size (FBS) was set to 5 as previous studies suggested [14].

To emphasize the imaging characteristics, image filters such as log (Laplacian of Gaussian), gradient, and lbp-2d/3d, and four common point-level transforms were applied prior to feature extraction. In addition, wavelet decomposition was applied at each channel for images to obtain all possible combinations in high-pass or low-pass filters (LLH, LHL, LHH, HLL, HLH, HHL, HHH, LLL). The radiomics features of each sequence image were extracted, and they included the following features: first-order features based on the pixel values of the images, shape features describing the morphology of the tumor, and a set of features describing the internal and surface textures of the lesion, Gray Level Co-occurrence Matrix (GLCM), Gray Level Run Length Matrix (GLRLM), Gray Level Zone Size Matrix (GLSZM), and Gray Level Dependence Matrix (GLDM). A total of 6420 radiomics features were extracted for each region of interest (ROI) in all the sequences and the z-score was normalized. The clinical variables (age, gender, tumor location, CEA status, and CA19-9 status) were subjected to statistical tests.

Dimensionality reduction and establishment of the model

Feature reduction, model construction, and external validation of the model were conducted based on a single sequence, including T1WI, T2WI, DWI, and CE-T1WI, and combined multiple sequences. Feature correlation analysis was first performed based on Pearson correlation analysis with a feature correlation threshold of 0.9. When the linear correlation coefficient between any two independent variables in the training set was greater than this threshold, one of the features was removed to alleviate redundancy between features. Priority was given to retaining the features with a higher linear correlation coefficient with the dependent variable. The tree-based approach was used for feature selection, where a tree model was built on the training data and impurity-based feature importance was used to select the important features and obtain the optimal feature subset. A random forest (RF) classifier was used to build the model. The internal validation of the model was completed using a 10-fold cross-validation method, which was a stronger method of model validation as the model was trained on 100% of the data. The generalization ability of the model was evaluated using an external validation set consisting of two hospitals. The ROC curves were used to assess model performance. The ROC curve is an indicator of the rate of false positives and true positives for continuous variables. The evaluation indicators mainly included AUC, sensitivity, specificity, and accuracy. The AUC was the main evaluation indicator of model performance.

Statistical analysis

SPSS (version 16.0) software was used for the statistical analysis of clinical parameters. For quantitative variables, the independent-samples t-test was used to detect differences between groups. For categorical variables, the X2 test was used to detect differences between groups. The Scikit learning package (version 0.20.3) was used to build the classification models. The ROC curves were plotted using Matplotlib (version 3.1.0).

Results

Clinical and laboratory characteristics

Based on the inclusion and exclusion criteria, 199 patients from three institutions were included in the study. There were 100 patients from hospital I who were used as the training set, including 36 MSI-H and 64 MS-L/S CRC patients. Furthermore, 99 patients from hospital II and hospital III were used as the validation set, including 32 MSI-H and 67 MS-L/S CRC patients. A total of 113 males and 86 females were included (average age: 57 years; range: 25–90 years).

The clinical and histopathological characteristics of rectal cancer patients included in our research are provided in Table 1. In both the training and external validation datasets, patient age, tumor location, gender, CA199, and CEA status were comparable between groups (p > 0.05). The clinical characteristics and laboratory features did not differ significantly between the training dataset and the validation dataset (p > 0.05).

Table 1 Characteristics of rectal cancer patients in the MS-L/S group and MSI-H group
Full size table

Predictive performance and external validation of the radiomics signature

After feature dimensionality reduction, the 20 most significant features were selected to construct the combined model based on multiple sequences of images. These included 6 features in DWI, 4 features in T2WI, 5 features in T1WI, and 5 features in CE-T1WI (Fig. 3).

Fig. 3
figure 3

Plot of feature importance F1: shape_Maximum2DDiameterSlice-CE-T1WI, F2: shape_MinorAxisLength-DWI, F3: shape_MinorAxisLength- CE-T1WI, F4: glrlm_GrayLevelNonUniformity-T2WI, F5: shape_Elongation-DWI, F6: gldm_DependenceEntropy-T1WI, F7: glcm_Idmn-T1WI, F8:original_glcm_Imc2-DWI, F9: glrlm_GrayLevelNonUniformity-T1WI, F10: gldm_LargeDependenceHighGrayLevelEmphasis-T2WI, F11: gldm_SmallDependenceLowGrayLevelEmphasis-DWI, F12: shape_Maximum2DDiameterSlice-DWI, F13: shape_Maximum2DDiameterColumn-DWI, F14: shape_SurfaceVolumeRatio-T1WI, F15:original_shape_Elongation- CE-T1WI, F16: gldm_SmallDependenceLowGrayLevelEmphasis-T2WI, F17: shape_Maximum2DDiameterRow- CE-T1WI, F18: shape_SurfaceVolumeRatio-T2WI, F19: glcm_ClusterShade- CE-T1WI, F20: glcm_ClusterProminence-T1WI

Full size image

In total, we obtained four models based on four independent sequences and a combined model based on multiple sequences images. The accuracy, sensitivity, specificity, and AUC value of the training set were listed in Tables 2 and 3. The CE-T1WI model had a comparable AUC value of 0.78 to the multi-sequence model, whose AUC value was higher than that of the remaining single sequence models (Table 3). After external validation, the AUC values of the T1WI model and T2WI model decreased significantly, from 0.74 to 0.67 and from 0.71 to 0.66, respectively, while the AUC values of the remaining models did not show a significant decline (Table 3). The ROC curves of the radiomics signature are shown in Fig. 4.

Table 2 Performance of the radiomics model in the training set and external validation set
Full size table
Table 3 The AUCs of the models in the training set and validation set
Full size table
Fig. 4
figure 4

The ROC curves of the radiomics signature in the training set and the external validation set

Full size image

Discussion

In this study, the clinical and laboratory characteristics were not included for model construction, as they were not statistically significant. Five radiomics models were established based on T1WI, T2WI, DWI, CE-T1WI, and multiple sequences (Fig. 4). The prediction models based on CE-T1WI and multiple sequences performed the best, with the highest AUC values and no significant decrease in the external validation set, suggesting that the developed predictive models were reliable and reproducible. They may be able to provide complementary information to determine MSI status in patients with rectal cancer to guide targeted therapy.

In clinical practice, MSI testing has the following important implications for CRC tumors: (1) MSI testing is used as a screening method for Lynch syndrome [15]; (2) MSI is a prognostic factor for stage II CRC [16]; (3) MSI is a predictor of the efficacy of adjuvant chemotherapy for stage II CRC [17]; (4) For the treatment of patients with stage IV CRC, MSI status helps to screen the population for whom the immune checkpoint inhibitor therapy is indicated. But it is important to note that there is a difference in MSI status between the colon and the rectum.

The frequency of MSI-H in a statistically linear fashion from the rectum to the ascending colon [18]. Given the low incidence of MSI rectal cancers, it is a challenge to study the effect of MSI on rectal cancer. Studies with a limited sample size predominate the literature. In Hasan et al’s study, only 8% of rectal cancers had their MSI status determined which may result in the selection bias issue, perhaps testing was reserved for cases that did not respond well to conventional chemotherapy, or perhaps it was only routinely being conducted at academic centers [3]. MSI has been consistently associated with both improved survival in colon cancer and decreased survival in rectal cancer [19, 20]. Locally advanced MSI rectal cancer was significantly associated with lower rates of pCR in response to neoadjuvant chemoradiotherapy [3]. Previous studies have shown the value of a CT image–based model for identifying the MSI status of CRC [21, 22]. It is well-known that MRI has a higher soft tissue resolution compared to CT, and it is the preferred imaging modality for rectal cancer [23]. It already has a key role in triaging patients for pre-operative chemoradiotherapy [24, 25]. it may be used as an imaging biomarker of MSI status.

As each sequence has a different role, it is difficult to fully characterize a tumor by relying on a single sequence alone. T1WI and T2WI provide basic morphological features of the tumor, while the inhomogeneity of DWI images reflects tumor heterogeneity [23, 26, 27]. Previous studies were based on only one or two non-enhanced sequences of MRI, whereas our research included all sequences of rectal MRI in the study. Compared to non-enhanced sequences, the model based on CE-T1WI was able to obtain an optimal predictive value compared to that of the model based on multiparametric images, with better generalization performance. This may be related to the histopathological characteristics of MSI-H tumors. MSI-H tumors are usually larger in size; tend to have mixed morphological features of glandular, mucinous, and solid components; and have a higher microvascular density [28, 29]. The enhancement sequence reflects not only the differences in tumor perfusion but also the differences in pixels in the mucinous and tumor cell-rich areas of the tumor.

The model developed by Zhang et al based on T2WI images accurately predicted the MSI status of rectal cancer with an AUC of 0.895 [12]. Li et al combined T2WI and ADC images to build the model, which had an AUC of 0.908 [30]. Although we expanded the sample size, the AUC values in the training set were still lower than those in their study. In our study, the AUC values of the model constructed based on T1WI and T2WI sequence images showed a significant decrease in the external validation set, demonstrating poor reproducibility of the model. Therefore, models derived from T1WI and T2WI sequences that have not been validated for repeatability may not be reliable. This is the problem of single-center research. It is difficult to compare and reproduce these results in studies at different institutions due to large differences in scan equipment and parameter settings, image pre-processing, the method of ROI segmentation, and the process of feature extraction [31]. The application of multi-center external validation provides a diverse range of imaging information to better assess the reproducibility of the model.

However, it should be emphasized that the model in this study has moderate predictive efficacy for the prediction of MSI status, and it may complement, but not replace histological MSI detection.

There are some limitations to our study. First, manual segmentation of the ROI used in this study is highly subjective and time-consuming, and further applications of semi-automatic or automatic segmentation are needed. Second, there were difficulties in outlining the lesions on T1WI, we referred to the enhanced images to determine the extent of lesions on T1WI. This may have implications for modelling, but the T1WI sequence was still included in this study to make it more comprehensive. Third, in order to retain as much image information as possible, all sequences were resampled to 1, 1, 1 mm isotropic voxels which may result in heavy interpolation. Fourth, although patients in our research were from three institutions, the number of patients with MSI-H rectal cancer was small; thus, the sample size of this study was relatively small. Further investigation with a larger sample size is still required.

In summary, the prediction and generalization abilities of the models based on CE-T1WI and multiparametric images are better than those of the model based on single unenhanced sequence images, which were moderately associated with the MSI mutation status of rectal cancer. However, this requires further validation with more samples before it can be used in clinical applications.