Prediction of Non-Response to Neoadjuvant Chemoradiotherapy in Esophageal Cancer Patients with 18F-FDG PET Radiomics Based Machine Learning Classification

Abstract

Background: Approximately 26% of esophageal cancer (EC) patients do not respond to neoadjuvant chemoradiotherapy (nCRT), emphasizing the need for pre-treatment selection. The aim of this study was to predict non-response using a radiomic model on baseline ¹⁸F-FDG PET.

Methods: Retrospectively, 143 ¹⁸F-FDG PET radiomic features were extracted from 199 EC patients (T1N1-3M0/T2-4aN0-3M0) treated between 2009 and 2019. Non-response (n = 57; 29%) was defined as Mandard Tumor Regression Grade 4-5 (n = 44; 22%) or interval progression (n = 13; 7%). Randomly, 139 patients (70%) were allocated to explore all combinations of 24 feature selection strategies and 6 classification methods towards the cross-validated average precision (AP). The predictive value of the best-performing model, i.e AP and area under the ROC curve analysis (AUC), was evaluated on an independent test subset of 60 patients (30%).

Results: The best performing model had an AP (mean ± SD) of 0.47 ± 0.06 on the training subset, achieved by a support vector machine classifier trained on five principal components of relevant clinical and radiomic features. The model was externally validated with an AP of 0.66 and an AUC of 0.67.

Conclusion: In the present study, the best-performing model on pre-treatment ¹⁸F-FDG PET radiomics and clinical features had a small clinical benefit to identify non-responders to nCRT in EC.

Keywords: esophageal neoplasms; neoadjuvant therapy; positron-emission tomography.

Figure 1

Inclusion and exclusion flowchart. Abbreviations: nCRT = neoadjuvant chemoradiotherapy.

Figure 2

Radiomics machine learning pipeline to train and select a model predicting non-response to nCRT. Radiomic and clinical features were normalized up front (blue area). Hyperparameter tuning was performed on the training subset (green area) with 24 unique feature selection strategies and 6 classification methods. The model with the highest mean average precision (AP) over the different cross validation folds was selected. The performance of this model was tested on the test subset (red area). Abbreviations: Skew = skewness of the distribution, SVM = support vector machine, NB = Gaussian Naive Bayes, KNN = K-nearest neighbors, RF = random forest, and NN = neural network.

Figure 3

Heatmap revealing radiomic feature clusters with similar expression (standardized on white-blue gradient scale) using unsupervised clustering with Pearson correlation as a measure of similarity. The x-axis represents the preselected radiomic features (n = 56) and the y-axis represents esophageal cancer patients in the training subset (n = 139). The heatmap reveals a substantial amount of feature redundancy.

Figure 4

Plot of the 10 best-performing models ordered by the mean average precision over the validation runs in the training subset (blue). The test performance was evaluated on an independent test set (red).

Figure 5

Precision–recall curve of the best performing model demonstrating the trade-off between precision and recall. The area under the precision–recall curve is reflected by the average precision. The average precisions for the training and test subset are 0.47 and 0.66, respectively. The black dashed line is the score of a random classification (0.28).

Figure 6

Learning curve of the best-performing model for prediction of non-response after nCRT in esophageal cancer. The average precision is plotted on the y-axis and the number of training samples on the x-axis. The training and test average precision scores did not fully converge to a point of stability yet, suggesting that the training process may slightly benefit from a larger sample size.