Predicting esophageal cancer response to neoadjuvant therapy with magnetic resonance imaging radiomics

Abstract

BACKGROUND

Predicting the pathological response of esophageal cancer (EC) to neoadjuvant therapy (NAT) is of significant clinical importance.

AIM

To evaluate the pathological response of NAT in EC patients using multiple machine learning algorithms based on magnetic resonance imaging (MRI) radiomics.

METHODS

This retrospective study included 132 patients with pathologically confirmed EC, were randomly divided into a training cohort (n = 92) and a validation cohort (n = 40) in a 7:3 ratio. 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). Radiomics signatures were selected using minimal redundancy maximal relevance and the least absolute shrinkage and selection operator. Nine classification algorithms were used to build the models, and the diagnostic performance of each model was evaluated using the area under the curve (AUC), sensitivity (SEN), and specificity (SPE).

RESULTS

A total of 1834 features were extracted. Following feature dimension reduction, ten radiomics features were selected to construct radiomics signatures. Among the nine classification algorithms, the ExtraTrees algorithm demonstrated the best diagnostic performance in both the training (AUC: 0.932; SEN: 0.906; SPE: 0.817) and validation cohorts (AUC: 0.900; SEN: 0.667; SPE: 0.700). The Delong test proved no significance in the diagnostic efficiency within these models (P > 0.05).

CONCLUSION

T2WI radiomics may aid in determining the pathological response to NAT in EC patients, serving as a noninvasive and quantitative tool to assist personalized treatment planning.

Key Words: Esophageal cancer; Neoadjuvant therapy; Pathological response; Magnetic resonance imaging; Radiomics

Core Tip: Few studies have utilized multiple radiomic algorithms to predict the pathological response to neoadjuvant therapy (NAT) in esophageal cancer (EC). In this study we found ten radiomics features related to pathological therapeutic response. The ExtraTrees algorithm performed good efficiency in the predictive and validation sets. Our study showed that the radiomics model derived from magnetic resonance imaging T2-weighted imaging images demonstrates good performance in determining the pathological response of NAT in EC, which would help individualized plans in EC.

INTRODUCTION

Esophageal cancer (EC) is a highly malignant digestive neoplasm, and radical resection combined with lymph node dissection is the standard treatment method[1,2]. However, a significant proportion of patients presents with local progression at diagnosis and are not candidates for surgery. Neoadjuvant therapy (NAT) prior to radical resection has been proved to reduce tumor size, stage, and surgical complexity, making it the first-line treatment option for patients with locally advanced EC[3]. However, some patients are resistant to NAT, thus receiving unnecessary treatment and experiencing toxic side effects, which negatively impact their prognosis[4]. Therefore, predicting the pathological response of EC to NAT at an early stage could help identify responders and reduce unnecessary treatment.

Currently, the assessment of treatment response to NAT in EC primarily relies on medical imaging techniques. Magnetic resonance imaging (MRI) could provide excellent anatomical details and soft tissue contrast, making it particularly useful for predicting tumor response after treatment[5-7]. According to the 2022 Chinese Society of Clinical Oncology guidelines, MRI has been recommended as a level 3 staging system for EC. The predictive value of using MRI to evaluate NAT response has achieved good results in various malignant tumors[8]. Recently, the development of artificial intelligence (AI) has further helped overcome the limitation of traditional imaging methods. The integration of AI with medical imaging has proven to be a powerful approach for enhancing the prediction of therapeutic response and prognosis, thereby aiding in the optimization of clinical treatment plans[9,10].

Radiomics, an advanced imaging analysis technique that extracts quantitative features from medical images, has gained widespread acceptance in clinical oncology for its ability to aid in predicting preoperative tumor staging and evaluating prognosis[11-13]. In particular, recent studies have demonstrated that radiomics features derived from imaging modalities can help predicting treatment response and prognosis in EC patients[14,15]. Despite the growing body of evidence supporting the usefulness of radiomics in EC, few studies have specifically addressed the use of MRI-based radiomic features in predicting NAT response. Among the MRI sequences, T2-weighted imaging (T2WI) has been less frequently explored for this purpose. T2WI, which provides high-resolution anatomical details, may offer distinct advantages in assessing tumor response to NAT, particularly with respect to changes in tumor microstructure and tissue composition that occur during treatment.

This study aimed to fill this gap by utilizing T2WI MRI-based radiomic features in conjunction with advanced machine learning algorithms to predict the pathological response to NAT in EC patients. The ultimate goal is to enhance the ability to stratify patients based on their likelihood of responding to NAT, thereby guiding more personalized treatment approaches and improving clinical outcomes.

MATERIALS AND METHODS

Patients

This retrospective study included 132 patients with pathologically confirmed EC who underwent NAT followed by radical surgery between January 2016 and October 2022. The inclusion criteria were: (1) Postoperative pathology confirmed EC; (2) Preoperative MRI scans performed within two weeks prior to NAT; (3) Clinical staging of II to IV; and (4) No history of other malignance. The exclusion criteria were as follows: (1) Any therapies administered prior to MRI examination; (2) Interruption of NAT for more than 2 weeks; (3) No surgical resection following NAT; (4) Poor quality of MRI images; and (5) Incomplete clinical data. All patients underwent radical EC resection surgery 4-6 weeks following completion of NAT. Based on the postoperative pathological remission status, patients were classified into a response group [tumor regression grades (TRGs) I to II] with better remission response and a non-response group (TRGs III to V) with poorer remission. This study was approved by the Ethics Committee on Clinical Researches and Novel Technologies of Meizhou People’s Hospital (grant No. 2021-CY-32). Written informed consent was waived by the ethics committee due to the retrospective nature of the study. All methods were performed in accordance with the relevant ethical guidelines and regulations.

Clinical and pathological data

Relevant clinical information was extracted from the hospital’s electronic medical record system, including patients’ age, gender, carcinoembryonic antigen (CEA), squamous cell carcinoma antigen (SCCA). The pathological results include tumor location, tumor size, degree of differentiation, TNM staging and tumor pathological response level.

Image acquisition

All examinations were conducted using a 3.0 T MRI system (Magnetom Skyra, Siemens Healthineers, Germany) equipped with an 18-channel body array coil. To minimize motion artifacts, patients were positioned supine and instructed to maintain quiet, regular breathing throughout the scan. The scanning range was from the level of the supraclavicular fossa to the gastric cardia. The scanning sequence and parameters were: (1) T1-weighted imaging (T1WI) Dixon sequence: Positive phase TR: 4.0 ms, TE: 1.3 ms, thickness: 5 mm, spacing: 6.5 mm; reverse phase TR: 4.0 ms, TE: 2.5 ms, thickness: 5 mm, spacing: 6.5 mm; (2) T2WI TSE BLADE sequence: TR: 7408 ms, TE: 84 ms, thickness: 5 mm, spacing: 6.5 mm; and (3) Diffusion-weighted imaging (DWI) sequence: TR: 6900 ms, TE: 59 ms, thickness: 5 mm, spacing: 6.5 mm, B values including 50 s/mm², 800 s/mm², and 1000 s/mm², respectively. The axial T2WI images were used to build the radiomics models. All image matrix was 210 × 210.

Pathological evaluation

The specimens excised after NAT were evaluated by a senior gastrointestinal pathologist. According to the standards of Mandard et al[16], tumor tissue was classified into TRG1 to TRG5: (1) TRG1: Complete remission, and tumor cells completely disappeared; (2) TRG2: A small amount of residual tumor cells with fibrotic proliferation; (3) TRG3: There are more residual tumor cells, but fibrosis and proliferation are still the main forms; (4) TRG4: The proportion of residual tumor cells is greater than that of fibrotic proliferation; and (5) TRG5: No fibrotic or necrotic reactions.

Image segmentation

DICOM raw data were imported into the 3Dslicer software. Two doctors with more than 10 years’ experience of abdominal imaging manually delineated and segmented the tumors along tumor margin on the axial T2WI images with unaware of the pathological results, and tumor volumes were generated on the software platform automatically. During the drawing and segmenting process, tumor necrosis, adjacent fat tissue and uninvolved esophagus wall should be excluded. Then another senior diagnostic doctor review and check all the region of interests (ROIs), with any discrepancies were resolved by discussion.

Image preprocessing

All segmented images were imported into the AK software (Artificial Intelligence Kit V3.3.0, GE Healthcare, China). The following steps were followed for image preprocessing according to previous reports[17,18]: (1) Step 1: Resampling of the images using B-spline interpolation to standardize the voxel size to 1 mm³; and (2) Step 2: Grayscale normalization, according to the Image Biomarker Standardization Initiative guidelines[19].

Radiomic feature extraction

Patients were randomly divided into a training cohort (n = 92) and a validation cohort (n = 40) at a 7:3 ratio, as according to previous report[20]. High-throughput extraction of radiomic features from the training cohort was executed with the Pyradiomics toolkit. The extracted feature set encompassed a comprehensive suite of 1834 metrics, including first-order statistics, morphological (shape-based) features, and second-order textural features derived from gray-level co-occurrence, run-length, size zone, dependence, and neighboring difference matrices. Subsequent to extraction, feature normalization was performed using the z-score technique (mean = 0, standard deviation = 1) to mitigate the influence of data scale variations across different feature types, consistent with conventional radiomics preprocessing practices[21]. Feature data extracted from the training cohort were used for feature selection, dimensionality reduction and prediction model constructions, while data from the validation cohort were used for validate the performance and effectiveness of the constructed models.

Feature selection and dimensionality reduction

Feature dimensionality reduction was performed to ensure the robustness of these models. Firstly, the intra-class correlation coefficients (ICCs) are calculated. Only features demonstrating excellent reproducibility (ICC > 0.75) were retained for subsequent analysis. t-test or Mann-Whitney U tests are then used to screen features with P < 0.05. Spearman’s rank coefficient was applied to determine the correlation between features with high repetitiveness. During this process, correlation coefficient between features was calculated as in reference to previous reports[22,23], and feature was retained with a correlation coefficient exceed 0.9. Following feature selection using the least absolute shrinkage and selection operator (LASSO) algorithm with 10-fold cross-validation, features with non-zero coefficients were retained to construct the radiomics signature. Nine classification algorithms, including logistic regression, support vector machines, K-nearest neighbor (KNN), random forest, ExtraTrees, extreme gradient boosting, light gradient boosting machine, multilayer perceptron and adaptive boosting, were applied to calculate a radiomics score for each patient. The area under curve (AUC) was calculated to assess model performance. The radiomics workflow was shown in Figure 1.

Figure 1 The workflow of radiomics analysis. T2WI: T2-weighted imaging.

Statistical analysis

We used Python (version 3.5.6) and SPSS 19.0 software to perform statistical analysis. Continuous variables were analyzed using independent sample t-tests (normal distribution) or Mann Whitney U tests (non-normal distribution), while categorical variables were analyzed using χ² tests or Fisher's exact probability test. ROC and AUC were used to evaluate the predictive performance of each model. Clinical application was evaluated by decision curve analysis (DCA), and the AUC of each model was compared by Delong test. A two-sided P value < 0.05 was considered statistically significant.

RESULTS

Patient characteristics

The clinical characteristics of all patients are summarized in Table 1. No statistically significant differences were observed in age, gender, or tumor location, degree of differentiation, length of tumor, or levels of CEA and SCCA between the response group and non-response groups (P > 0.05).

Table 1 Clinical characteristics of all esophageal cancer patients.

Characteristics	Non-response group	Response group	P value
Age (year)	60.38 ± 8.96	55.44 ± 7.99	0.347
Gender, n (%)			0.375
Male	78 (81.25)	31 (86.11)
Female	18 (18.75)	5 (13.89)
Location, n (%)			0.765
Upper	17 (17.71)	4 (11.11)
Middle	46 (47.92)	20 (55.56)
Lower	23 (23.95)	9 (25.00)
Middle and lower	10 (10.42)	3 (8.33)
Degree of differentiation, n (%)			0.388
Well differentiated	1 (1.04)	1 (2.78)
Moderately differentiated	74 (77.08)	29 (80.56)
Poorly differentiated	14 (14.59)	5 (13.88)
Medium to high differentiation	1 (1.04)	0 (0)
Middle to low differentiation	6 (6.25)	1 (2.78)
Length of tumor	5.54 ± 1.76	5.43 ± 1.91	0.862
CEA	3.09 (1.78, 3.26)	2.89 (1.74, 3.65)	0.47
SCCA	2.40 (0.80, 3.00)	1.70 (0.74, 1.86)	0.859

CEA: Carcinoembryonic antigen; SCCA: Squamous cell carcinoma antigen.

Radiomics features

A total of 1834 radiomics features were extracted. After feature selection and dimensionality reduction, ten radiomics features were retained for NAT response in EC. The names and descriptions of the selected features are provided in Figure 2.

Figure 2 The names and descriptions of the selected features.

Radiomics signature discrimination performance

The radiomics signature was evaluated using nine different classification algorithms. The AUC values in the training cohort ranged from 0.735 to 1.000, and in the validation cohort from 0.733 to 0.900 (Table 2). Among these, the ExtraTrees algorithm demonstrated the best diagnostic performance in both the training [AUC: 0.932; sensitivity (SEN): 0.906; specificity (SPE): 0.817] and validation cohort (AUC: 0.900; SEN: 0.667; SPE: 0.700). The Delong test showed no significant difference in diagnostic efficacy among these algorithms (P > 0.05; Table 3). The receiver operating characteristic curves for each model in both cohorts are shown in Figure 3. DCA further indicated that the ExtraTrees model provided the highest net benefit across a range of probability thresholds (Figure 4).

Figure 3 Performance evaluation of radiomics models. A and B: Receiver operating characteristic curve analysis comparing model performance in the training (A) and validation cohort (B). AUC: Area under the curve; KNN: K-nearest neighbor; GBM: Light gradient boosting machine; LR: Logistic regression; MLP: Multilayer perceptron; SVM: Support vector machines; XGBoost: Extreme gradient boosting.

Figure 4 Decision curve analysis demonstrating clinical utility across probability thresholds in the validation cohort. The Y-axis indicates net benefit standardized by patient population. The blue curve corresponds to the radiomics model, the black solid curve represents the all-patients-treated scenario, and the black dashed line indicates the no-patients-treated scenario. DCA: Decision curve analysis.

Table 2 Performance of different classification algorithms in predicting neoadjuvant therapy efficacy in esophageal cancer.

Models	Task	AUC	95%CI	Sensitivity	Specificity	Accuracy
LR	Train	0.798	0.7127-0.8841	0.844	0.634	0.693
LR	Test	0.800	0.5144-1.0000	0.667	0.600	0.615
SVM	Train	0.735	0.6310-0.8393	0.781	0.598	0.649
SVM	Test	0.733	0.3511-1.0000	0.333	0.800	0.692
KNN	Train	0.848	0.7814-0.9148	0.375	0.915	0.763
KNN	Test	0.783	0.4851-1.0000	0.667	0.700	0.692
RF	Train	0.962	0.9321-0.9910	0.906	0.854	0.868
RF	Test	0.833	0.5562-1.0000	0.667	0.600	0.615
ET	Train	0.932	0.8832-0.9811	0.906	0.817	0.842
ET	Test	0.900	0.6801-1.0000	0.667	0.700	0.692
XGBoost	Train	1.000	1.0000-1.0000	0.969	1.000	0.991
XGBoost	Test	0.767	0.4999-1.0000	0.667	0.700	0.692
LGBM	Train	1.000	1.0000-1.0000	0.969	1.000	0.991
LGBM	Test	0.800	0.5507-1.0000	0.667	0.700	0.692
AdaBoost	Train	1.000	1.0000-1.0000	0.969	1.000	0.991
AdaBoost	Test	0.800	0.5203-1.0000	0.667	0.600	0.615
MLP	Train	0.808	0.7179-0.8987	0.781	0.805	0.798
MLP	Test	0.767	0.4938-1.0000	0.667	0.700	0.692

AUC: Area under the curve; LR: Logistic regression; SVM: Support vector machines; KNN: K-nearest neighbor; RF: Random forest; ET: ExtraTrees; XGBoost: Extreme gradient boosting; LGBM: Light gradient boosting machine; MLP: Multilayer perceptron; AdaBoost: Adaptive boosting.

Table 3 Delong test for comparing area under the curves among classification algorithms.

Models	LR	SVM	KNN	RF	ET	XGBoost	LGBM	AdaBoost	MLP
LR		0.405	0.900	0.480	0.194	0.794	1.000	1.000	0.820
SVM	0.405		0.729	0.194	0.110	0.852	0.648	0.546	0.863
KNN	0.900	0.729		0.637	0.095	0.893	0.895	0.859	0.873
RF	0.480	0.194	0.637		0.230	0.582	0.710	0.626	0.648
ET	0.194	0.110	0.095	0.230		0.230	0.275	0.160	0.246
XGBoost	0.794	0.852	0.893	0.582	0.230		0.480	0.745	1.000
LGBM	1.000	0.648	0.895	0.710	0.275	0.480		1.000	0.777
AdaBoost	1.000	0.546	0.859	0.626	0.160	0.745	1.000		0.794
MLP	0.820	0.820	0.873	0.648	0.246	1.000	0.777	0.794

LR: Logistic regression; SVM: Support vector machines; KNN: K-nearest neighbor; RF: Random forest; ET: ExtraTrees; XGBoost: Extreme gradient boosting; LGBM: Light gradient boosting machine; AdaBoost: Adaptive boosting; MLP: Multilayer perceptron.

DISCUSSION

This study introduced a radiomics framework designed to predict personalized responses to NAT in individuals with EC. A key distinction from prior investigations is the utilization of the T2WI sequence from MRI to derive radiomics features. The findings suggested that MRI-based approach had potential as a valuable instrument for clinicians to detect non-responsive patients early in the course of NAT, thereby playing an assisting role in refining the comprehensive treatment strategy for appropriate EC cases.

EC prognosis varies significantly due to tumor heterogeneity. While radiomics methods using enhanced computed tomography (CT) or 18F-FDG positron emission tomography (PET) images have been employed to assess NAT response[24-26], few studies have explored magnetic resonance radiomics for this purpose. Hou et al[27] extracted 138 radiomics features from T2WI images of 68 EC patients, demonstrating strong predictive performance for radio-chemotherapy response. Compared with conventional CT or PET-CT, T2WI offers superior lesion detail visualization, avoids radiation exposure, and eliminates concerns related to renal impairment or contrast agent allergies[28]. Therefore, we leveraged T2WI images and a machine learning approach to develop an optimal radiomics model for predicting NAT response in EC patients. Ten radiomics features were selected from the T2WI sequences. These features primarily reflect differences in pixel grayscale intensity and morphological patterns. Radiomics features characterize the spatial distribution patterns of pixels (voxels) within the tumor. Greater tumor heterogeneity typically manifests as more irregular voxel distributions, which may correlate with treatment response and patient prognosis[29]. Results from our study demonstrate that MRI T2WI radiomics demonstrated good performance in predicting the pathological response to NAT in EC patients. Our findings may assist in guiding individualized treatment and care for EC patients, potentially improving prognosis.

Various methods can be used to build radiomics signatures, which can result in different prediction accuracy. In the present study, we established nine classification algorithms to build radiomics models based on T2WI images. By calculating the key features in these algorithms, the ExtraTrees algorithm exhibited the best diagnostic accuracy and stability in the primary cohort and also good performance in the validation set. This may be contributed to the advantage of ExtraTrees algorithm in avoiding over-fitting ability for high-dimensional data, and its robustness, stability and good generalization ability[30]. Our results indicate that the radiomics algorithms based on T2WI images could predict pathological response of EC with high accuracy, and is consistent with previous reports[27]. Results from DCA indicate the ExtraTrees model exhibit good diagnostic efficiency, which is consistence with previous results of Hou et al[27] and Lu et al[31]. These models identify tumor characteristics beyond visual observation, and utilize machine learning algorithms revealing comprehensive situation of tumor heterogeneity.

Radiomics features that extracted from MRI images could provide more detailed variations in tissue composition and tumor heterogeneity[32,33]. Hu et al[34] predicted an AUC of 0.852 for the efficacy of radiotherapy and chemotherapy in EC based on the CT radiomics characteristics of tumors. Yang et al[35] predicted the efficacy of neoadjuvant radiotherapy and chemotherapy for EC based on the radiomics features selected by LASSO. The results showed that its AUC in the training cohort was 0.84-0.86, and in the testing cohort it was 0.71-0.79. Kawahara et al[36] developed a multi-institution prediction model based on PET/CT radiomics and dosimetry features. The mixed model achieved accuracy rates of 84.4%, 86.0% and 79.0% for CT radiomics based on KNN classifier, PET radiomics based on PET classifier, and dosomics based on KNN classifier, respectively. Our MRI-based radiomics model demonstrated superior performance compared to the aforementioned CT-based models, suggesting the potential added value of MRI in characterizing EC tumor heterogeneity.

No clinical factors were found to be significantly associated with pathological response to NAT in EC in this study, which is consistent with most previous reports[37,38]. However, Chao et al[39] identified age as a risk factor. This discrepancy may be attributed to differences in sample size or the characteristics of the cohort populations included in these studies.

This research has several limitations. First, its single-center origin and retrospective nature, coupled with a limited sample size, mean that our findings must be validated in future multicenter studies with larger, prospective cohorts. Second, manual delineation of the ROI may introduce inter-observer variability, potentially affecting the consistency of the results. Last, we incorporated only a limited set of clinical factors. Future studies should integrate a broader range of clinical features with radiomics characteristics which including DWI and contrast-enhanced T1WI to better elucidate the relationship between these factors and the pathological response to NAT in EC.

CONCLUSION

This study demonstrates that MRI T2WI-based radiomics can effectively predict pathological response to NAT in EC. The T2WI-derived models may serve as a valuable clinical tool to guide personalized treatment planning for EC patients.