A radiomics model based on diffusion-weighted imaging developed using machine learning enables prediction of microsatellite instability in endometrial cancer

Abstract

Background

The objective of this research was to develop and validate a machine learning-based prediction model integrating clinical data, apparent diffusion coefficient (ADC) value, and diffusion- weighted imaging (DWI)-based radiomic features, aimed at assessing microsatellite instability (MSI) status in endometrial cancer (EC) patients.

Methods

In total, 292 EC patients who underwent pelvic MRI scans participated in this study and were allocated into three distinct groups: an external validation cohort (n = 70), a testing cohort (n = 68), and a training cohort (n = 154). Preoperative clinical indicators, ADC metrics, and radiomic parameters extracted from DWI images were comprehensively evaluated. Feature selection was conducted using least absolute shrinkage and selection operator (LASSO) regression combined with Mann-Whitney U testing. Following feature selection, three distinct machine learning classifiers—support vector machine (SVM), random forest (RF), and logistic regression (LR)—were employed to develop predictive models. The performance and clinical utility of these models were subsequently examined through receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis (DCA).

Results

Among the evaluated methods, the RF model incorporating two clinical indicators, six radiomic parameters from DWI, and ADC values exhibited superior predictive ability. The areas under the ROC curves (AUC) reached 0.980 (95% CI: 0.944–0.996) in the training cohort, 0.852 (95% CI: 0.745–0.927) in the test cohort, and 0.938 (95% CI: 0.853–0.982) in the external validation cohort, respectively. These AUC values reflected better accuracy compared to separate predictive models employing only clinical factors, DWI radiomics, or ADC values (training cohort: AUC = 0.682, 0.925, and 0.851; test cohort: AUC = 0.663, 0.788, and 0.731; external validation cohort: AUC = 0.605, 0.872, and 0.828, respectively). Calibration curves indicated robust concordance, and DCA analysis confirmed that the model had substantial clinical applicability.

Conclusion

A prediction model combining clinical factors, DWI radiomics features, and ADC values with machine learning algorithms can noninvasively assess MSI status in EC.

Introduction

Endometrial cancer (EC) represents one of the most critical gynecological malignancies, with steadily increasing global rates of incidence and mortality [1]. Microsatellite instability (MSI), which is characterized by defects in DNA mismatch repair (MMR) mechanisms and resultant hypermutability, is observed in approximately 30% of EC patients [2]. In a meta-analysis, Xiao et al. systematically reviewed the survival data of patients with endometrioid EC [3]. The analysis found that microsatellite instability (MSI) may correlate with improved or distinct survival outcomes, and highlighted MSI’s role as a predictive biomarker for immunotherapy response. Evrard et al. also confirm that MSI status is a key prognostic marker of EC [4]. Nonetheless, conventional MSI assessment techniques—including polymerase chain reaction (PCR), next-generation sequencing (NGS), and immunohistochemistry (IHC)—are costly, invasive, and laborious, imposing economic burdens on patients [5,6,7]. Therefore, it is of paramount clinical significance to establish a reliable, economical, and minimally invasive approach to detect MSI in EC.

Diffusion-weighted imaging (DWI), a widely adopted quantitative magnetic resonance imaging (MRI) technique in clinical imaging, provides enhanced morphological information compared to conventional MRI sequences [8], delineating lesion localization and multiplicity clearly, and quantitatively assesses restricted diffusion of tissue water molecules through apparent diffusion coefficient (ADC) measurements. Consequently, DWI offers considerable promise for accurately characterizing microscopic lesion features [9]. Earlier research has demonstrated the utility of DWI and ADC metrics in evaluating various histopathological attributes of EC, including histological classification, tumor grading, proliferative status, and genetic alterations [10,11,12]. Prior investigations, however, predominantly focused on lower-order imaging parameters, such as morphological assessments and ADC measurements, due to methodological limitations. Consequently, high-dimensional radiomic features extracted from DWI data remain largely unexplored, restricting potential advances in diagnostic accuracy. Radiomics is a promising and rapidly evolving discipline that quantitatively extracts substantial, high-dimensional data from medical imaging, identifying potential associations with clinical outcomes. It has become increasingly valuable for decision-making in clinical practice [13, 14]. Recent radiomics research in EC has primarily concentrated on predicting lymph node metastasis LNM, lymphovascular space invasion, and patient risk stratification [15,16,17]. Nonetheless, the use of radiomics analysis for MSI determination in EC remains limited. Previous radiomics investigations primarily utilized computed tomography (CT) or conventional MRI sequences, with fewer reports addressing radiomics features derived specifically from DWI [18,19,20].

Therefore, this study sought to design and verify an innovative, non-invasive prediction model integrating clinical parameters, ADC values, and high-dimensional radiomic features from DWI, utilizing advanced machine learning methods. This model aims to offer an accurate and efficient approach for preoperative MSI assessment in patients diagnosed with EC.

Materials and methods

Ethics and study participants

This research was approved by the Institutional Review Board of the First Affiliated Hospital of Xinxiang Medical University. Initially, from July 2016 to January 2025, a total of 346 individuals suspected of EC from center I and 125 from center II, all of whom had undergone pelvic MRI examinations, were recruited. Inclusion criteria are as follows: (1) Diagnosed with EC through surgery resection or tissue biopsy; (2) No contraindications for MRI examination, such as metallic implants in the pelvic cavity or claustrophobia. Exclusion criteria encompassed the following conditions: (1) MSI testing was unavailable because of financial or other factors (n = 110); (2) chemotherapy or radiation therapy had been administered prior to MRI (n = 24); (3) either missing DWI sequences or poor-quality DWI data unsuitable for extracting radiomic features (n = 17); and (4) preoperative clinical data was incomplete (n = 28). Consequently, the final cohort comprised 222 patients from center I and 70 from center II. Recorded clinical parameters included age, tumor dimensions (largest diameter on axial T2WI), and tumor marker concentrations, such as carcinoembryonic antigen (CEA), carbohydrate antigen 199 (CA 199), carbohydrate antigen 125 (CA 125), and carbohydrate antigen 153 (CA 153).

MSI status assessment

The MSI status was determined via IHC staining targeting four key MMR proteins, namely mutL homolog 1, mutS homolog 2, mutS homolog 6, and PMS1 homolog 2. For EC tissue samples, those showing intact expression of all four MMR proteins were classified as MSS. In contrast, EC tissues exhibiting reduced or absent expression of one or more of these MMR proteins were defined as MSI [21]. All IHC staining analyses and MSI/MSS classification were independently performed by two pathologists with extensive clinical experience. In cases of conflicting judgments, discrepancies were addressed and resolved through joint consultation between the two experts.

MRI protocol

MRI examinations were performed at both centers using 3.0 T scanners (Signa HDxt; GE Healthcare, Milwaukee, WI, USA). Imaging was performed with patients in a supine position, spanning from the pubic symphysis up to the anterior superior iliac spine. The DWI parameters are as follows: repetition time/echo time = 3437 / 71.4 ms, flip angle = 90°, slice thickness/slice interval = 5 / 1 mm, matrix = 256 × 260, and b = 800 s/mm².

Tumor segmentation

Tumor segmentation was manually performed on axial DWI slices without referencing any prior histopathological or clinical results, using axial T2WI sequences for visual guidance. Initially, a radiologist (reader 1) with six years of experience outlined the tumor contours slice-by-slice using ITK-SNAP software (v3.8.0). Then, another experienced radiologist (reader 2), who had accumulated 15 years of practice, independently reviewed, corrected, and confirmed the delineations. Ultimately, volumes of interest (VOIs) of entire tumor were produced by aggregating regions of interest (ROIs) from each tumor slice through the software’s three-dimensional function.

ADC value calculation

The VOIs generated from DWI sequences were then projected onto their corresponding ADC maps, generating ADC values based on the entire tumour. ADC values were computed based on the following formula: S_b / S₀ = exp (- b × ADC), in which S_b and S₀ represent signal intensities at specified b-values and a b-value of 0, respectively, and b denotes the diffusion sensitizing factor.

Feature extraction

Radiomic features were derived from the DWI sequences by employing PyRadiomics software (version 2.1.2) (Fig. 1). Prior to feature extraction, images underwent N4 bias field correction, z-score normalization, voxel size resampling at 1 mm³, and were discretized using a bin width of 20 [22]. Twelve filters (Original, BinomialBlurImage, Laplacian Sharpening, AdditiveGaussian Noise, BoxMean, CurvatureFlow, Wavelet, ShotNoise, LoG, Normalize, DiscreteGaussian, and BoxSigmaImage) were subsequently applied. As a result, 2264 radiomic parameters were extracted per DWI image.

Fig. 1

Flow chart of the study

Feature selection

Radiomic feature reproducibility was evaluated based on DWI scans from 20 randomly chosen patients, focusing on inter- and intra-observer agreement. Reader 1 performed a second segmentation after a two-week interval to assess intra-observer consistency, whereas reader 2 separately conducted VOI segmentation and feature extraction to examine inter-observer consistency. Initially, only radiomic features displaying intra- and inter-observer intraclass correlation coefficients (ICCs) exceeding 0.75 were selected [18]. Subsequently, the Mann–Whitney U test was employed to remove radiomic features that lacked significant discriminatory power (P < 0.05) between MSI and MSS tumors. The retained features underwent z-score normalization, after which redundant features were eliminated via least absolute shrinkage and selection operator (LASSO) regression analysis (regularization parameter α set to 0.001, Fig. 2). Clinical parameters, limited in number, were directly screened using the Mann–Whitney U test, while ADC values were included directly into the predictive models without additional feature selection.

Fig. 2

Radiomics feature selection using the LASSO. (a) LASSO coefficient profile plot; (b) LASSO parameter plot

Fig. 3

SHAP analysis of combined (clinical + DWI + ADC, red) model

Model development

Cases from center I were randomly split into training and test cohorts at a 7:3 ratio, and cases from center II served as an external validation group. To optimize model performance, variables selected for modeling were scaled using the Min-max normalization method prior to constructing predictive models. Three machine learning classifiers, random forest (RF), logistic regression (LR), and support vector machine (SVM), were individually employed to build distinct predictive models based separately on clinical characteristics, ADC values, and DWI radiomic features, as well as a combined integrated model (clinical + ADC + DWI). Consequently, a total of 12 prediction models were generated (4 models × 3 classifiers). Parameters set for the LR algorithm included a penalty (L2), a penalty factor (C) of 1.0, no class weighting, and a tolerance level of 0.0001. For the RF model, parameters included the Gini criterion, a minimum of two samples per split, one sample per leaf as minimum, and 100 estimators. For the SVM classifier, parameters included the RBF kernel, a gamma value of 0.01, a penalty factor (C) of 1.0, and a classification threshold of 0.5. To mitigate overfitting and enhance robustness, the finalized predictive model was required to exhibit an AUC difference of less than 0.15 between training and test datasets, alongside the highest AUC performance in the test cohort [23].

Statistical analysis

Statistical analyses were performed using SPSS version 23.0 and R software (v3.5.3). Continuous variables were analyzed using the Mann–Whitney U test, whereas categorical variables were analyzed by Chi-square tests; statistical significance was set at a P-value of less than 0.05. Model diagnostic performance was quantified by the the area under the receiver operating characteristic (ROC) curve (AUC), and DeLong’s test was applied for comparison. To further evaluate incremental improvements by the comprehensive model, net reclassification index (NRI) and integrated discrimination improvement (IDI) metrics were computed. Calibration curves were generated to assess prediction consistency, and decision curve analysis (DCA) was conducted to evaluate clinical utility and overall net benefit of the models.

Result

Clinicopathologic data and ADC value

The training dataset comprised 154 patients (56 MSI and 98 MSS), the test dataset included 68 patients (24 MSI and 44 MSS), and the external validation dataset consisted of 70 patients (25 MSI and 45 MSS). Across the three cohorts (training, test, and external validation), ADC values were significantly lower among MSI patients (P < 0.001, P = 0.044, and P < 0.001, respectively) (Table 1).

Table 1 Summary of characteristics in differents sets

Prediction model

Regarding individual predictive models, an LR-based clinical model incorporating two clinicopathologic markers (CEA and CA199) was selected. The RF algorithm was applied to the DWI model, ultimately including six radiomic features(wavelet-glcm-wavelet-LHL- ClusterProminence, wavelet-glcm-wavelet-LHL-Imc2, wavelet-glcm-wavelet-HLH-Idmn, additive Gaussiannoise-glszm-Small AreaHigh GrayLevel Emphasis, normalize-glszm- SmallAreaEmphasis, and wavelet-gldm-wavelet-HH-Dependence Variance). The ADC predictive model utilized the SVM algorithm. For the combined model (clinical + DWI + ADC), the final chosen configuration was RF-based, integrating the two clinical indicators (CEA and CA199), the six DWI radiomic parameters, and ADC measurements (Table 2, Fig. 3). Additionally, during combined model construction, we assessed multicollinearity among the aforementioned nine features. The results indicated their variance inflation factors were 1.176, 1.172, 1.227, 2.024, 2.062, 1.227, 1.151, 1.250, and 1.187, respectively.

Table 2 Predictive performance of different models

Prediction performance

Among all predictive models, the comprehensive approach incorporating clinical characteristics, ADC measurements, and DWI radiomic features achieved optimal diagnostic performance, displaying AUCs of 0.980, 0.852, and 0.938 in the training, test, and external validation cohorts, respectively. In the training group, this integrated approach significantly outperformed models constructed solely with clinical parameters (AUC = 0.682, Z = 6.877, P < 0.001), ADC metrics (AUC = 0.851, Z = 3.457, P = 0.001) and DWI-based radiomics model (AUC = 0.925, Z = 2.399, P = 0.016). Regarding the test cohort, the combined predictive method showed significantly improved AUC compared with clinical (AUC = 0.663, Z = 2.349, P = 0.019) and ADC-based models (AUC = 0.828, Z = 2.425, P = 0.019), but its advantage over the DWI model (AUC = 0.788, Z = 1.209, P = 0.227) was not statistically significant. In external validation, the combined predictive method showed significantly improved AUC compared with clinical (AUC = 0.605, Z = 5.563, P < 0.001) and ADC-based models (AUC = 0.828, Z = 2.091, P = 0.037), but its advantage over the DWI model (AUC = 0.872, Z = 1.497, P = 0.134) was not statistically significant (Table 2; Fig. 4).

Fig. 4

ROC curves of clinical (blue), DWI (green), ADC (yellow), and combined (clinical + DWI + ADC, red) models for predicting MSI status. (a) Training set ROC curves; (b) test set ROC curves; (c) external validation set ROC curves

Furthermore, the combined predictive model (clinical + DWI + ADC) demonstrated improved accuracy in reclassifying MSI status risk. In the training cohort, the calculated NRIs relative to the clinical, DWI, and ADC models were 154.08%, 28.06%, and 63.27%, respectively; in the test cohort, these figures were 88.64%, 38.64%, and 60.61%; in the external validation cohort, 133.33%, 82.67%, and 110.22%, respectively. Similarly, IDIs were enhanced in the training set (73.47%, 17.56%, 33.33%), the test set (35.95%, 15.59%, 18.89%), and the external validation set (59.40%, 17.48%, 13.33%) (Table 3).

Table 3 The NIR and IDI of Clinical, DWI, and ADC models

Model validation and clinical utility

Calibration curves revealed strong concordance between predictions derived from the combined model (clinical + DWI + ADC) and observed outcomes across all datasets (training, test, external validation), and the Hosmer-Lemeshow test yielded chi-square values of 2.162, 6.386, and 5.666, with corresponding p-values of 0.976, 0.604, and 0.685 respectively. (Fig. 4). DCA confirmed that the combined model provided consistent and robust clinical benefit in predicting MSI status in EC patients across the three sets (Figs. 5 and 6).

Fig. 5

Calibration curves of clinical (red), DWI (green), ADC (orange), and combined (clinical + DWI + ADC, blue) models. (a) Training set calibration curve; (b) test set calibration curve; (c) external validation set calibration curve

Fig. 6

DCA for the combined model (clinical + DWI + ADC, red). (a) Training set; (b) test set; (c) external validation set

Discussion

With the advancement of computer technology, radiomics and machine learning approaches have demonstrated great potential in the assessment of oncology, especially EC. For instance, a meta analysis by Gao et al. concluded that a predictive model based on MRI radiomic features could accurately assess the risk stratification of EC [24]. Additionally, Ying et al. demonstrated that machine learning approaches were effective in evaluating the MSI status of gastric cancer [25]. Meanwhile, Broomand et al. and Capello et al. also revealed that predictive models built on radiomic features—whether derived from MRI or CT—could effectively evaluate the MSI status in EC patients [26, 27]. Inspired by the above research, we developed a predictive model combining two clinical indicators, six radiomic parameters derived from DWI, and ADC measurements to predict MSI status among EC patients. Compared with individual models based solely on clinical variables, DWI radiomic features, or ADC values, the integrated model exhibited enhanced diagnostic performance and improved risk stratification, underscoring its potential as a valuable clinical decision-making tool in EC treatment planning.

Currently, clinical applications of DWI primarily involve lesion localization through changes in signal intensity and microscopic evaluation using ADC values [28]. Previous studies exploring the value of DWI for MSI assessment showed that MSI tumors had more malignant and compact characteristics than MSS tumors, leading to increased restriction of water molecule diffusion, higher DWI signal intensity, and lower ADC values [7, 29]. In our study, ADC values in the MSI group were significantly lower than those in the MSS group across the training, test, and external validation sets (AUC = 0.851 and 0.731, respectively). These findings align with earlier studies, confirming the feasibility of using DWI to assess MSI status in EC. However, in clinical practice, differences in DWI signal intensity between MSI and MSS groups are often challenging to distinguish visually, and ADC values overlap, limiting accurate MSI assessment using conventional DWI methods.

To maximize the utility of information embedded in DWI scans and improve MSI prediction accuracy, we utilized machine learning techniques to establish three distinct predictive models: clinical-only, DWI-only, and a combined model integrating clinical parameters, DWI-derived radiomics, and ADC values. Within the clinical model, despite evaluating numerous candidate factors, only CEA and CA199 emerged as significant predictors in the final selection. From a clinical perspective, CEA and CA199 are tumour markers associated with abnormal cell proliferation and metabolic dysregulation. MSI in EC arises from defects in the DNA mismatch repair system, which drive hypermutability and aggressive tumour phenotypes. Elevated CEA/CA199 levels may therefore reflect this heightened oncogenic activity. However, the diagnostic performance of clinical models is relatively low (AUC values: training = 0.682, test = 0.663, validation = 0.605). This limited performance might reflect the absence of robust, clinically validated biomarkers specific to MSI, which reduces the predictive strength of standard clinical metrics [30, 31]. In contrast, the DWI-based model consisted of six selected radiomic features, demonstrating promising predictive capability (AUC values: training = 0.925, test = 0.788, validation = 0.872). Above six radiomic features capture microstructural and textural heterogeneity. For instance, wavelet-based features reflect spatial variations in tumour cell density and stromal organisation, while GLCM-based features quantify tissue texture, both of which are linked to MSI-related pathological changes (e.g. irregular cell nesting and increased inflammatory infiltration). These features convert microscopic biological heterogeneity into quantifiable imaging signals, thereby bridging the gap between molecular MSI status and macroscopic imaging phenotypes. These outcomes corroborate prior radiomics-based studies employing traditional MRI, supporting the effectiveness of DWI-derived radiomics alone in determining MSI status in EC [18, 19]. Furthermore, the combined predictive model, incorporating clinical features, DWI radiomics, and ADC values, further augmented the diagnostic accuracy and improved risk classification relative to the individual models. Such a result aligns with earlier findings, emphasizing the superiority of integrated multidimensional models in comprehensively characterizing tumors and accurately predicting MSI status [19, 32, 33].

The RF classifier, an ensemble learning approach based on the bagging algorithm, offers advantages including high predictive precision, robustness against overfitting, and effectiveness in handling both regression and classification tasks [34]. LR, a generalized linear modeling technique, facilitates accurate estimation of variable effects and effectively captures underlying associations among multiple variables [35]. The SVM classifier aims to reduce classification errors in previously unseen data without assumptions regarding the underlying probability distribution, thus excelling at detecting subtle and intricate patterns within complex datasets [36]. These three algorithms have been widely applied for evaluating various diseases. For example, Yu et al. demonstrated that an MRI radiomics model based on RF could predict axillary lymph node metastasis in breast cancer [37]; Sun et al. showed that a CT radiomics model using LR effectively distinguished benign from malignant bone tumors [38]; and Wu et al. revealed that an MRI radiomics model based on SVM could diagnose Parkinson’s disease [39]. In this study, all three algorithms were included in the modeling process. The optimal algorithm differed across models: LR was optimal for the clinical model, RF was optimal for the DWI and combined models, and both LR and SVM were optimal for the DWI model. This variation may result from different optimal application scenarios among machine learning algorithms. Thus, future trends in clinical practice might involve using multiple algorithms to select the best predictive model [40].

This model demonstrates certain application potential in the pre-treatment assessment of patients with EC. Clinicians may utilize this model to attempt noninvasive prediction of MSI status through pelvic MRI data (DWI sequences and ADC values) as well as basic clinical indicators (CEA and CA199). If subsequent validation proves feasible, this approach may help in formulating treatment plans and avoid reliance on costly invasive tests such as PCR or NGS. The model does not require complex radiomics pipelines: its core radiomics features can be extracted and selected using the PyRadiomics tool, LASSO, and Mann-Whitney U test, respectively. Technically, this process can be automated via analysis software. Theoretically, clinicians only need to input standard MRI data and clinical parameters, and the system can directly output MSI prediction results. However, the actual implementation effect of this workflow still requires further verification. In conclusion, although the model demonstrate promising performance, substantial additional validation studies are necessary before their utility in real-world practice can be confidently established.

This study has several limitations. Firstly, the sample size of 292 patients with small MSI subgroups lacks statistical power and is at risk of overfitting. Secondly, patients from only two centres introduced geographical homogeneity and selection bias, resulting in a lack of broad representativeness. Thirdly, manual tumour segmentation introduces subjectivity, which could inflate the discriminative power of features and hinder standardisation. Fourth, while many machine learning algorithms exist, utilising only LR, RF and SVM in this study may result in more effective models being overlooked. Fifth, as the study aimed to explore prediction models based on DWI radiomic features, conventional MRI sequences such as T1WI and T2WI were not included. This may limit the comprehensiveness of lesion assessment. In the future, we will conduct prospective multicenter studies involving larger and more diverse patient cohorts and try to automate lesion segmentation while incorporating additional imaging sequences and a broader range of machine learning and deep learning algorithms with the aim of achieving more stable, reliable, and precise research outcomes.

Conclusion

Collectively, a machine learning-based predictive model integrating clinical parameters, DWI radiomic features, and ADC metrics may provide a non-invasive and reliable tool for assessing MSI status in EC patients, potentially assisting clinicians in therapeutic decision-making.