Magnetic resonance imaging-based radiomics signature for predicting preoperative staging of esophageal cancer

Abstract

BACKGROUND

Esophageal cancer (EC) is one of the most prevalent malignant gastrointestinal tumors; accurate prediction of EC staging has high significance before treatment.

AIM

To explore a rational radiomic approach for predicting preoperative staging of EC based on magnetic resonance imaging (MRI).

METHODS

This retrospective study included 210 patients with pathologically confirmed EC, randomly divided into a primary cohort (n = 147) and a validation cohort (n = 63) in a ratio of 7:3. All patients underwent a preoperative MRI scan from the neck to the abdomen. High-throughput and quantitative radiomics features were extracted from T2-weighted imaging (T2WI) and gadolinium contrast-enhanced T1-weighted imaging (T1WI)-Gd images. Radiomics signatures were selected using minimal redundancy maximal relevance and the least absolute shrinkage and selection operator. Then a logistic regression model was built to predict the EC stages. The diagnostic performance of the radiomics model for discriminating between stages I-II and III-IV was evaluated using the area under the curve (AUC), sensitivity (SEN), and specificity (SPE).

RESULTS

A total of 214 radiomics features were extracted. Following feature dimension reduction, the T1WI and T2WI sequences were retained, and 14 features from the T1WI sequence and 3 features from the T2WI sequence were selected to construct radiomics signatures. The radiomics signature combining T2WI with T1WI-Gd demonstrated superior discrimination of stages in the validation cohort (AUC: 0.851; SEN: 0.697; SPE: 0.793), which outperformed single-sequence models (AUC: 0.779, 0.844; SEN: 0.667, 0.636; SPE: 0.8, 0.8).

CONCLUSION

MRI-based radiomics signatures could identify EC stages before treatment, which could serve as a noninvasive and quantitative approach aiding personalized treatment planning.

Key Words: Esophageal cancer; Tumor staging; Magnetic resonance imaging; Radiomics; Logistic regression

Core Tip: This study developed a novel magnetic resonance imaging-based radiomics approach to noninvasively predict preoperative stage of esophageal cancer (EC). By integrating quantitative features from T2-weighted imaging and contrast-enhanced T1-weighted imaging sequences in 210 EC patients, a logistic regression model achieved high accuracy in distinguishing early-stage (I-II) from advanced-stage (III-IV) disease. This multimodal radiomics signature outperformed single-sequence models and offers a promising tool for guiding personalized treatment strategies.

INTRODUCTION

Esophageal cancer (EC) poses a significant health burden in China, ranking sixth in incidence and fourth in cancer mortality, with a dismal 5-year survival of 19%[1,2]. Early-stage (I-II) disease typically undergoes endoscopic or surgical resection, whereas advanced (III-IV) EC requires neoadjuvant chemoradiation prior to surgery or definitive chemoradiation[3]. Lymph node (LN) metastasis correlates strongly with reduced survival[4], underscoring the imperative for accurate pretreatment staging. The accurate prediction of EC staging prior to treatment is clinically critical.

Radiomics is increasingly recognized as a critical tool in cancer research. High-throughput mining enables the extraction of quantitative image features from digitally encoded medical images. These robust image-based signatures could potentially enhance precision diagnosis and treatment strategies[5,6]. Recent radiomics research has demonstrated the possibility of using magnetic resonance imaging (MRI) to substantially improve the detection and prediction of LN metastases[7,8]. While computed tomography (CT), MRI, positron emission tomography-CT (PET-CT), and endoscopic ultrasonography (EUS) each exhibit variable sensitivity (SEN)/specificity (SPE) for EC staging and LN detection[9-11], contrast-enhanced CT remains the clinical mainstay despite controversial LN detection accuracy (SEN: 37.3%-67.2%; SPE: 63.9%-96.4%)[12]. Notably, the new Chinese EC guidelines has now recognized MRI’s diagnostic parity with CT for LN assessment[13].

At present, radiomics research on EC based on MRI mainly focuses on predicting LN metastases and evaluating the efficacy of neoadjuvant therapy[14-16], However, there has been little research on predicting the early and local progression stages of EC based on MRI radiomics, which has significant clinical significance for personalized treatment of EC patients. Our study addresses this by extracting radiomic features from T2-weighted imaging (T2WI) and T1-weighted imaging (T1WI)-Gd sequences to build a predictive model for distinguishing early (I-II) from locally advanced (III-IV) EC, providing a reference for guiding clinical treatment decisions.

MATERIALS AND METHODS

Patients

The clinical and MRI data of EC patients who underwent LN dissection confirmed by surgery and pathology between March 2017 and November 2020 were retrospectively collected. The inclusion criteria were: (1) Pathologically confirmed EC; (2) Underwent radical resection and LN dissection, with clear postoperative pathological stage and clear pathological LN results; (3) Preoperative MRI scan (T2WI and T1WI-Gd) within two weeks before surgery; (4) No anticancer treatment prior to MRI; and (5) No history of other malignant tumors. The exclusion criteria were: (1) Received preoperative anticancer therapy; (2) Concurred other malignant tumors; and (3) Non-diagnostic MRI image quality, precluding subsequent feature extraction.

This study was approved by the Ethics Committee on Clinical Researches and Novel Technologies of Meizhou People’s Hospital (Grant No. 2023-C-45), and patient informed consent was waived for this retrospective study. All procedures were carried out in accordance with relevant guidelines and regulations. Ultimately, 210 patients with EC (165 males and 45 females, aged 40 to 86 years, mean age 63.1 ± 7.92 years) were included. Among them, there were 207 cases of squamous cell carcinoma, 2 cases of adenosquamous cell carcinoma, and 1 case of mixed neuroendocrine carcinoma. Based on tumor node metastasis staging, 101 patients had stages I-II (74 males, 27 females, mean age 63.2 ± 8.01 years), and 109 had stages III-IV EC (91 males and 18 females, mean age 62.9 ± 7.86 years). The clinicopathologic characteristics of the patients are presented in Table 1.

Table 1 Clinical and pathological characteristics of 210 esophageal cancer patients.

Characteristics	Primary cohort		P value	Validation cohort		P value
	Stage I-II	Stage III-IV		Stage I-II	Stage III-IV
Age (years)	62.0	61.0	0.548	62.5	64.0	0.842
Gender			0.073			0.613
Male	50	63		24	28
Female	21	13		6	5
Location			0.33			0.705
Upper	2	6		2	2
Middle	47	51		17	22
Lower	22	19		11	9

Image acquisition

Imaging was performed with a 3.0 T Siemens (Erlangen, Germany) Skyra MR scanner equipped with 18-channel phased array coil. The patient was placed supine and told to breathe as calmly as possible to reduce motion artifacts. The scanning range was from the level of the supraclavicular fossa to the gastric cardia. The scanning sequence and parameters were: (1) T2WI: Slice thickness 3 mm, TR 3000 ms, TE 91 ms, NEX = 1, matrix 256 × 256, field of view (FOV) 210 mm × 210 mm and scanning time 2 minutes; and (2) T1WI-Gd: Slice thickness 1.5 mm, TR 4.4 ms, TE 2.0 ms, NEX = 1, matrix 160 × 160, FOV 150 mm × 150 mm, and flip angle 12. During enhanced scanning, Gd-DTPA contrast agent was injected intravenously through the elbow using a high-pressure syringe at 0.1 mmol/kg. The scanning time was 3 minutes and 26 seconds.

Tumor segmentation

Manual segmentation of the EC was performed on each patient's MR images utilizing an Artificial Intelligent Kit (AK), a new application platform for the overall solution of artificial intelligence in medical imaging radiomics launched by GE Healthcare. Two experienced associate senior radiologists (readers 1 and 2) carefully delineated the tumor on all T2WI and T1WI-Gd images to generate two 3D segmentations of the entire tumor. These regions of interests were applied for subsequent feature extraction and further analysis.

Radiomic feature extraction

Patients were randomly divided into a primary cohort (n = 147) and a validation cohort (n = 63) at a 7:3 ratio, as according to previous reports[17,18], and the sample size was estimated by using the package of R software (version 4.3.1). Radiomic features were extracted based on the segmentation results from the primary cohort using a homemade program in the AK. Preprocessing included: (1) The image was first resampled using linear interpolation algorithm to a standard voxel spacing of 1 mm × 1 mm × 1 mm; and (2) Z-score for MRI intensity values were standardized. The Python PyRadiomics module was used to extract radiomic features through AK, where the bin width of the histogram was set to 5 to quantify the intensity distribution of the image. The distance parameter for texture features such as gray level co-occurrence matrix (GLCM) was uniformly set to 1. For each MRI sequence, the 214 extracted features covered the major feature pool in recent radiomic studies. The features were categorized into six primary types (Figure 1): (1) Histogram features; (2) Form factor features; (3) GLCM and Harilick features; (4) Gray level run length matrix features; (5) Grey level size zone matrix (GLSZM) features; and (6) Gray level dependence matrix (GLDM) features.

Figure 1 Flowchart of the study. A: Original magnetic resonance imaging images, including T2-weighted imaging (T2WI) and T1-weighted imaging (T1WI)-Gd; B: Segmentation of both T2WI and T1WI-Gd images delineated the tumor region; C: Radiomic features were extracted from the tumor region, including shape, first-order histogram, texture and wavelet group analysis; D: A series of features were selected to build the radiomics signature using the minimal redundancy maximal relevance and the least absolute shrinkage and selection operator method; E: The classification performance of the radiomic signature was evaluated using the receiver operating characteristic curves in both the primary and validation cohort. ROC: Receiver operating characteristic; T1WI: T1-weighted imaging; T2WI: T2-weighted imaging.

A rigorous feature selection pipeline was implemented to simplify the model development and avoid over-fitting. Minimal redundancy maximal relevance (mRMR) and least absolute shrinkage and selection operator (LASSO) method were used for feature dimensionality reduction. For tuning coefficients λ and α, the respective minimum standard deviation and maximum area under the curve (AUC) criteria were followed. Dimensionality reduction was performed using the variance method (excluding features with zero variance), the mRMR method selecting the top 50 features, and the LASSO method with lambda selection standard set to ‘λ.min’, and selecting the optimal λ through five fold cross validation on the training set. For accurate prediction of the EC stages, 16 features selected from the T2WI with T1WI-Gd images made the greatest contribution. Then, each patient’s two radiomic signature sets were integrated into a multivariate logistic regression framework. Model training generated weighted linear combinations of selected features, which constituted the final radiomic signature scores for individual patients.

Statistical analysis

The statistical analysis was performed in R (version 3.5.1; http://www.Rproject.org). The Shapiro-Wilk test was used to test whether the quantitative data were normally distributed. The rank-sum test was used to compare the differences in age, between patients with EC stage I-II and stage III-IV. The χ² test or correction χ² test was used to compare the differences in sex, tumor location, and pathological type between patients with EC stage I-II and stage III-IV. In addition, the primary cohort was used to train the logistic regression model, and the validation cohort was used to test the model. Finally, the receiver operating characteristic (ROC) curve, SEN, and SPE were used to evaluate the performance of the prediction model. The DeLong test compared the AUC between the single-sequence and mixed-sequence models. A two-sided P value < 0.05 was considered statistically significant.

RESULTS

Clinical data

Patient clinicopathological features are summarized in Table 1. No significant intergroup differences (stages I-II vs III-IV) existed in age, sex, or tumor location (P > 0.05).

Radiomic features

After feature selection and dimensionality reduction, 14 and 3 of the 214 radiomic features were retained for EC staging prediction (Table 2). Fourteen and three features distinguished stages I-II from stages III-IV in the T1WI and T2WI models, respectively.

Table 2 Selected radiomics features of esophageal cancer patients.

Task	MRI	Feature name	Coefficient
Tumor node metastasis stage prediction	T2WI
		Original_firstorder_Median	0.921642536
		Original_shape_Elongation	0.967002183
		Original_glcm_Correlation	0.246781324
	T1WI-Gd
		Original_shape_LeastAxisLength	0.875999143
		Original_shape_Elongation	0.967002356
		Original_glcm_DifferenceAverage	0.109732763
		Original_firstorder_Kurtosis	0.605404872
		Original_firstorder_Median	0.921642536
		Original_glszm_LowGrayLevelZoneEmphasis	0.274812737
		Original_glcm_Imc1	0.638235237
		Original_gldm_LargeDependenceHighGrayLevelEmphasis	0.209360630
		Original_shape_MajorAxisLength	0.149464274
		Original_ngtdm_Strength	0.214134941
		Original_glcm_Idn	0.983044251
		Original_firstorder_10Percentile	0.303070619
		Original_glszm_SizeZoneNonUniformity	0.126542445
		Original_ngtdm_Complexity	0.524606933

MRI: Magnetic resonance imaging; T1WI: T1-weighted imaging; T2WI: T2-weighted imaging.

Radiomics signature discrimination performance

The radiomics signature was constructed using the selected features from the previous section, forming a linear combination based on a logistic regression model. There were four first-order features, four shape features, four GLCM features, two GLSZM features, one GLDM feature, and two neighboring gray-tone difference matrix features. The formula for the combined model was as follows: Radscore = T1WI × 5.27600436 + T2WI × 3.21937331 - 4.36897257.

The discriminative power of the radiomic signature for tumor, node, metastasis staging was assessed using ROC curves in both the primary and validation cohorts (Figure 2). Subsequently, the ROC curves corresponding to different models were compared using the Delong test, and the AUC, SEN, and SPE were calculated (Tables 3 and 4). The radiomics signature combining T2WI with T1WI-Gd exhibited superior discrimination of the stages in the primary (AUC: 0.788, 0.859; SEN: 0.75, 0.829; SPE: 0.746, 0.789) and validation cohorts (AUC: 0.851; SEN: 0.697; SPE: 0.793), outperforming single-sequence models (AUC: 0.779, 0.844; SEN: 0.667, 0.636; SPE: 0.8, 0.8). Radiomic scores (Rad-scores) for EC patients in both the primary and validation cohorts were calculated using the elastic net model.

Figure 2 Receiver operating characteristic curves of the radiomics signature’s discriminative performance in esophageal cancer staging in the primary and validation cohorts. A: Primary cohorts; B: Validation cohorts. The combined radiomics signature demonstrated superior area under the curve (AUC) in both the primary (AUC: 0.877) and validation cohorts (AUC: 0.851), outperforming single-sequence models. AUC: Area under the curve; T1WI: T1-weighted imaging; T2WI: T2-weighted imaging.

Table 3 Delong test results in the primary and validation cohort.

Model	Primary cohort			Validation cohort
	T2WI	T1WI	Combined	T2WI	T1WI	Combined
T2WI	0	0.081	0.0035	0	0.27	0.09
T1WI	-1.74	0	0.19	-1.09	0	0.80
Combined	-2.95	-1.30	0	-1.65	-0.25	0

T1WI: T1-weighted imaging; T2WI: T2-weighted imaging.

Table 4 Model performance for predicting tumor node metastasis stages in esophageal cancer patients.

Model	Primary cohort			Validation cohorts
	AUC (95%CI)	SEN (95%CI)	SPE (95%CI)	AUC (95%CI)	SEN (95%CI)	SPE (95%CI)
T2WI	0.788 (0.712, 0.858)	0.75 (0.642, 0.834)	0.746 (0.634, 0.833)	0.779 (0.651, 0.892)	0.667 (0.496, 0.803)	0.8 (0.627, 0.905)
T1WI	0.859 (0.799, 0.919)	0.829 (0.729, 0.897)	0.789 (0.681, 0.868)	0.844 (0.728, 0.941)	0.636 (0.466, 0.778)	0.8 (0.627, 0.905)
Combined	0.877 (0.819, 0.929)	0.842 (0.744, 0.907)	0.761 (0.65, 0.845)	0.851 (0.739, 0.948)	0.697 (0.527, 0.826)	0.8 (0.627, 0.905)

95%CI: 95% confidence interval; T1WI: T1-weighted imaging; T2WI: T2-weighted imaging; AUC: Area under the curve; SEN: Sensitivity; SPE: Specificity.

DISCUSSION

This study demonstrated that a quantitative approach utilizing multimodal MRI can effectively identify EC stages prior to treatment initiation. Leveraging significant differences in the Rad-score, the radiomics signature successfully distinguished between esophageal squamous cell carcinoma stages I-II and III-IV during the preoperative period. These findings were supported by MRI’s superior soft-tissue contrast and multiparametric capabilities, which provide high-quality imaging foundation for precise tumor delineation and radiomics analysis.

Surgical resection remains the cornerstone of EC management, yet surgical intervention for advanced EC (stages III-IV) is associated with significant challenges and a high postoperative recurrence rate. LN status is a critical independent prognostic factor, and the approach for LN dissection plays a pivotal role in treatment strategies[4]. Several studies have suggested that systematic LN dissection should be performed in EC patients only when the tumor invades the submucosa, as extended LN resection may elevate the risk of postoperative complications[19,20]. In addition, accurate LN staging is essential for guiding the implementation of neoadjuvant therapy[21]. Although CT is widely used for preoperative EC staging, its accuracy remains limited. Radiomics addresses these limitations by decoding tumor heterogeneity through high-throughput feature extraction from medical images, thereby enhancing diagnostic and prognostic precision.

This study confirmed the feasibility of predicting EC stages using MR-based radiomic features. We extracted 214 quantitative features from tumors using T2WI and T1WI-Gd images. The seventeen features that contributed most to the differentiation of EC stages were selected for model construction. Radiomics features derived from T2WI are based on the signal intensity of water molecules, primarily reflecting the water content characteristics of the tumor.

The high-order radiomics features comprised the GLCM, which reflects the gray level difference, including distribution uniformity and texture thickness. Greater heterogeneity of the tumor was associated with an uneven gray level and increased texture complexity. Histological features derived from T1WI-Gd images predominantly indicative of microvascular density and histologic grading[22]. Higher-order texture features, exemplified by GLCM metrics, quantified spatial complexity patterns correlating with biologically aggressive phenotypes[23].

Consistent with prior studies, our results showed that integrating radiomics features from dual-sequence MRI achieved strong discriminatory performance in differentiating stage I-II from stage III-IV EC in both the primary and validation cohorts[16,24,25]. Our MRI-based radiomics model outperformed CT and PET-CT radiomics methods for staging EC[26,27]. This advantage may originate from MRI's superior soft-tissue resolution coupled with synergistic information integration from multiparametric sequences. The improvement may also stem from the complementary information provided by multiparametric MRI, which enables a more comprehensive and objective assessment of tumor heterogeneity. In addition, our results demonstrated performance comparable with prior MRI radiomics studies predicting LN metastasis in EC[16]. This reinforces the potential of MRI-based radiomics for EC staging.

This study had several limitations. First, the single-center design and modest sample size necessitate validation through multicenter studies with larger cohorts. Second, although we focused on tumor staging, the prediction efficiency of LN metastasis, including its extent, in different regions remains unassessed. Future work will be dedicated to developing models focused on predicting LN metastasis. Last, the exclusive use of a single MRI scanner limits generalizability. Future work should evaluate the impact of varying equipment, field strength, and scanning parameters.

CONCLUSION

This study demonstrates that MRI-based radiomics can effectively differentiate between stage I–II and stage III-IV EC before treatment. We developed a multiparametric radiomics model incorporating features from multiple MRI sequences. MRI-based radiomics could serve as a valuable noninvasive adjunct tool for preoperative staging and personalized therapeutic strategy formulation in EC.