2.1 Study cohort and dataset

This retrospective study came from Shantou University Medical College Cancer Hospital. We conducted this study after obtaining approval from the institutional medical ethics committee (ethics number: 2023116). Since this study was retrospectively designed and data collection was based on existing medical records, written informed consent was waived. The study initially screened 175 patients who underwent surgical treatment for primary esophageal cancer between 2015 and 2022. All patients completed enhanced CT scans before surgery, and the expression of the immunohistochemical marker cytokeratin 19 (CK19) in tumor tissue was evaluated by pathological examination after surgery. The inclusion criteria for patients were: (1) a clear diagnosis of primary esophageal cancer by pathology; (2) complete preoperative CT imaging data; (3) complete postoperative pathological and immunohistochemical analysis reports. Exclusion criteria included: (1) patients who had received radiotherapy or chemotherapy before surgery; (2) CT imaging data quality was not sufficient to meet the needs of radiomics analysis; (3) patients with missing pathological results. The samples in this study represent the patient population with typical esophageal cancer characteristics in this institution, providing a reliable data basis for model development. The patient enrollment process is shown in Fig 1, and 134 patients were finally included in the follow-up study.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Flow chart of patient enrollment in this study.

https://doi.org/10.1371/journal.pone.0350744.g001

2.2 Imaging protocol

Preoperative CT scans in this study were performed using two CT scanners: a GE BrightSpeed 16-row scanner (GE HealthCare, Boston, USA) and a Philips Ingenuity 64-row scanner (Philips HealthCare, Amsterdam, Netherlands). The scan range was from the hypopharynx to the lower edge of the diaphragm to ensure full coverage of the esophagus and its adjacent structures. The scans were performed using spiral CT technology, with the slice thickness and interslice spacing set to 5 mm, the tube current automatically adjusted within 160–300 mA, and the tube voltage fixed at 120 kVp. The rotation time was 0.75 s per revolution, the pitch was set to 0.999, and the single scan time was approximately 7.5 s.

Before scanning, patients fasted for at least 6 hours and drank 600 mL of water in divided doses within 15–20 minutes before the examination to increase gastrointestinal contrast; an additional 200 mL of water was drunk immediately before the scan. During scanning, patients were positioned supine with their hands raised and crossed above the head and were fixed with an elastic band to reduce motion artifacts. All patients were instructed to take a deep breath and hold it to complete the scan in a single breath-hold.

Enhanced scans were performed using a bolus-tracking technique. A nonionic contrast agent (iopromide, 300 mg I/mL; Ultravist, Bayer, Germany) was injected intravenously at a rate of 2.5 mL/s, with a dose of 1.5 mL per kilogram of body weight (typically 75–100 mL), followed by a 30 mL saline flush. The region of interest (ROI) was placed in the thoracic aorta, and scanning was automatically triggered when the attenuation reached 150 HU, with a 3-second delay after the trigger to ensure optimal vascular enhancement and lesion contrast.

2.3 Tumor segmentation

This study selected venous phase enhanced CT images as the basis for manual segmentation of the region of interest (ROI) because it performs best in contrast enhancement between tumors and surrounding tissues and can significantly improve the recognition of tumor boundaries. All segmentation work is based on the original image files in DICOM format and is completed through the dedicated medical image processing software 3D Slicer. During the segmentation process, the largest lesion region (Largest Tumor Region, LTR) of the tumor is marked first. For patients with a single tumor, the complete lesion outline is directly outlined; for patients with multiple lesions, the largest lesion region that is representative and available for subsequent analysis is selected based on the size and location of the lesion. The main purpose of this strategy is to ensure that the segmented region represents the overall characteristics of the lesion to the greatest extent possible, while reducing the impact of multi-lesion segmentation on the complexity of the subsequent model.

To ensure the accuracy and consistency of the segmentation results, this study adopted a two-person blind segmentation process. The initial segmentation was completed by a radiologist with 5 years of imaging diagnosis experience, and the segmentation results were mainly based on the clarity of the tumor boundary on the image and the contrast characteristics with the surrounding tissue. Subsequently, another senior radiologist with 20 years of rich experience independently reviewed the initial segmentation results and made necessary adjustments to the segmentation boundaries based on anatomical characteristics, imaging characteristics, and his own clinical experience. Finally, all segmentation results were finalized through consensus discussion between the two doctors to eliminate possible subjective biases in the segmentation process. To adapt to subsequent imaging omics analysis, all segmented ROIs were saved as standardized mask files and registered with the original CT images to ensure data consistency and traceability.

2.4 Feature extraction and screening

In order to extract reliable radiomics features from esophageal cancer CT images, this study used Pyradiomics, a feature extraction tool widely used in medical image analysis. For each preprocessed CT image, we extracted features covering multiple aspects such as shape, texture, and signal intensity, including first-order statistics, shape features, gray-level co-occurrence matrix (GLCM), gray-level size region matrix (GLRLM), gray-level dependency matrix (GLDM), and gray-level run length matrix (GLSZM). These multi-dimensional imaging features provide a rich data foundation for subsequent analysis.

The feature screening process includes several key steps. The intraclass correlation coefficient (ICC) is used to evaluate feature stability. Features with an ICC value of more than 0.75 are considered stable and can be included in subsequent analysis. Features with high ICC values are Z-score standardized to eliminate dimensional and order of magnitude differences between features. Features significantly associated with disease status were screened by t-test, and features with p values less than 0.05 were statistically significant. The Lasso regression model was used to streamline the feature range, and L1 regularization was used to complete feature selection, simplify the model, and improve prediction accuracy. The regularization parameters were optimized through cross-validation to screen the features that contributed most to the prediction model. These features are used for machine learning model training to build an efficient and accurate preoperative non-invasive prediction model for CK19 expression.

2.5 Machine learning prediction model

In order to accurately predict CK19 markers non-invasively before surgery, this study used four mainstream machine learning models, including gradient boosting decision tree (GBDT), random forest (RF), extreme gradient boosting (XGB), and lightweight gradient boosting machine (LGBM). Each model was optimized and validated under a series of predefined parameters.

2.6 Statistical analysis

In this study, a variety of machine learning models were comprehensively evaluated and compared, aiming to develop an efficient model that can non-invasively predict CK19 marker expression before surgery. The experimental data were randomly divided into training and test sets in a ratio of 7:3. The training set was used for model construction and parameter optimization, while the test set was used to independently verify the predictive performance of the model to ensure the generalization ability of the model on unknown data. For the performance evaluation of the model, multiple indicators such as accuracy, recall, precision, F1 score, and area under the receiver operating characteristic curve (AUC) were used. In addition, the specific performance of the model under different classification categories was analyzed in detail through the confusion matrix to reveal potential misclassification patterns. In order to further explain the prediction mechanism of the model and improve its clinical interpretability, the SHAP (SHapley Additive exPlanations) value analysis tool was introduced. Based on the principles of game theory, SHAP can quantify the contribution of each feature to the model output. In this exploratory analysis, performance metrics are reported as point estimates without bootstrap confidence intervals or formal statistical comparisons, consistent with the hypothesis‑generating nature of the study.

3.1 Experimental setup

The experiments in this study were completed on a high-performance computing workstation running the Windows 10 operating system, equipped with an Intel Core i9 processor and an NVIDIA GeForce RTX 3080 graphics card to ensure computational efficiency for model training and validation. All data processing and analysis used the Python 3.7 programming language, combined with a variety of open source libraries, including Scikit-learn, XGBoost, LightGBM, Pandas, and Numpy, to automate the entire process of data preprocessing, feature extraction, model training, and performance evaluation. In the experiment, radiomics features were extracted from standardized CT image masks, and the data were divided into training and test sets (7:3 ratio) under stratified random sampling to ensure that the proportion of CK19-positive and -negative cases in different groups was consistent to reduce the impact of sample distribution bias on the model.

3.2 Feature extraction and screening results

In this study, a total of 993 radiomics features were extracted from preoperative enhanced CT images of 134 patients, including 320 features from the original images and 673 features generated by wavelet transform. The original image features encompass multiple families, including shape, first‑order statistics, gray‑level co‑occurrence matrix (GLCM), gray‑level run length matrix (GLRLM), gray‑level size zone matrix (GLSZM), gray‑level dependence matrix (GLDM), and neighborhood gray‑tone difference matrix (NGTDM). The wavelet‑transformed features similarly cover these families across different frequency subbands, capturing tumor information at multiple scales and frequency domains.

In order to reduce feature redundancy and screen the features that contribute most to model prediction, this study uses the Lasso regression method. By introducing the L1 regularization constraint, Lasso regression can compress the weights of unimportant features to zero, achieve feature sparsification and optimize model performance. After preliminary screening, a total of 72 features showed significant contributions, and then the regularization parameters were further optimized through cross-validation, and finally 11 most critical features were screened out. Fig 2a shows the Lasso coefficient ranking of these key features, among which wavelet- HLL_glszm_GrayLevelNonUniformityNormalized, logarithm_glszm_LargeAreaHighGrayLevelEmphasis, exponential_ngtdm_Complexity and other features have a significant influence in the prediction model. Fig 2b shows the change of Lasso coefficient with regularization strength (α value). As the α value increases, most feature coefficients gradually tend to zero, and only a small number of features that contribute significantly to the model are retained.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. (a) Shows the key radiomics features selected by Lasso regression and their contributions, and (b) describes the dynamic changes of feature coefficients under changes in regularization strength (α value).

https://doi.org/10.1371/journal.pone.0350744.g002

3.3 Preoperative non-invasive prediction of CK19 expression results

This study compared the performance of four mainstream machine learning models (GBDT, RF, XGB, and LGBM) in preoperative noninvasive prediction of CK19 expression, and the results are shown in Table 1. Among the four models, random forest (RF) showed relatively higher accuracy and precision, with an accuracy of 0.8293, a precision of 0.6824, a recall of 0.6134, an F1 score of 0.6325, and an AUC value of 0.6765. The gradient boosted decision tree (GBDT) was slightly inferior to RF in accuracy (0.7805) and recall (0.6408), but its F1 score (0.6328) was slightly higher, showing a better balanced performance. The performance of extreme gradient boosting (XGB) and lightweight gradient boosting machine (LGBM) was relatively close, both showing lower precision and recall (both 0.5167 and 0.5126), and AUC values of 0.6744 and 0.6807, respectively.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Prediction results of various machine learning models in the validation set.

https://doi.org/10.1371/journal.pone.0350744.t001

Given the class imbalance in the dataset (83% CK19-negative vs. 17% CK19-positive), we additionally report balanced accuracy and Matthews correlation coefficient (MCC) to provide a more robust evaluation. For the random forest model, balanced accuracy was 0.613 and MCC was 0.288, indicating modest discriminative ability that is consistent with the AUC value.

For the specific performance analysis of the RF model (as shown in Fig 3), the model calibration curve (Fig 3a) shows good consistency between the predicted probability and the true probability. Although there is a certain deviation in the low probability interval, the overall trend is close to the perfect calibration curve. The ROC curve (Fig 3b) shows the classification ability of the model at different decision thresholds, and the AUC value is 0.6765, indicating that the RF model has certain advantages in distinguishing positive and negative samples. Decision curve analysis (Fig 3c) shows that the RF model yields a net benefit within a limited range of threshold probabilities; however, for a substantial portion of the threshold range, net benefit falls below zero. This indicates that clinical net benefit is confined to specific risk thresholds, and the model’s clinical applicability requires further validation in larger cohorts.. The confusion matrix (Fig 3d) shows that the RF model can accurately identify most negative samples (32/34), but there are certain deficiencies in the prediction of positive samples (2/7), suggesting that the model still has room for optimization in positive example recognition.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. (a) The model calibration curve shows the comparison between the predicted probability and the actual probability, which is used to evaluate the calibration performance of the model.

(b) The ROC curve describes the classification ability of the model at different decision thresholds, and the AUC value reflects the overall performance of the model; (c) The decision curve (Net Benefit graph) evaluates the clinical application benefits of the model at different risk thresholds; (d) The confusion matrix intuitively displays the classification results of the optimal model on the test set, including the distribution of true positive, false positive, true negative and false negative.

https://doi.org/10.1371/journal.pone.0350744.g003

Overall, the RF model showed modest predictive performance in this exploratory study, providing reliable technical support for preoperative noninvasive prediction of CK19 expression. However, its recall rate and AUC value still have room for improvement, and future studies can try to further improve the model performance through feature optimization or ensemble learning methods.

3.4 SHAP visualization analysis

To further reveal the prediction mechanism of the random forest (RF) model and improve its clinical interpretability, this study used SHAP (SHapley Additive exPlanations) values to analyze and visualize the important features of the model. Based on the principle of game theory, the SHAP value can quantify the impact of each radiomics feature on the model prediction output, and use positive and negative values to indicate the direction of its action in driving the model output to increase or decrease. Fig 4 shows the SHAP value distribution of the RF model, which lists the 11 radiomics features that contribute most to the prediction results. The horizontal axis is the SHAP value corresponding to the feature, indicating the degree of influence of the feature on the prediction output; the vertical axis is the specific feature name. The color of the point represents the size of the feature value, red indicates a higher feature value, and blue indicates a lower feature value.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Visual analysis of the SHAP values of the RF model, showing the impact of important radiomics features on the model prediction results and their direction.

https://doi.org/10.1371/journal.pone.0350744.g004

As can be seen from the figure, wavelet- HLL_firstorder_Median and exponential_firstorder_Uniformity are the two features that have the most significant impact on the model prediction, with a high SHAP value distribution. This indicates that the statistical characteristics of the tumor region, such as texture consistency and median, play a key role in distinguishing CK19 positive from negative expression. In addition, the features wavelet- LLH_gldm_DependenceNonUniformityNormalized and logarithm_glszm_LargeAreaHighGrayLevelEmphasis also show important contributions, which are mainly related to the density distribution and grayscale size area uniformity of the tumor region. It can be further observed from the color distribution that the high values (red) of some features (such as wavelet- HLL_firstorder_Median) mainly push the predicted value to move in the positive direction, suggesting that it has a higher feature value in samples with CK19 positive expression. The low values (blue) of another part of the features (such as exponential_firstorder_Uniformity) also have a positive impact on the prediction results, indicating that the reduction of these feature values may be associated with CK19 positivity.