1469P - Development of an efficacy prediction model for concurrent chemoradiotherapy in esophageal squamous cell carcinoma using deep learning and multimodal data integration

Background

Employing deep learning techniques, this study aims to construct a dual-center predictive model for evaluating the efficacy of concurrent chemoradiotherapy in esophageal squamous cell carcinoma. Through the amalgamation of pre-treatment CT images, clinical features, dosimetric parameters, radiomic features, and 3D dose distribution data, a comprehensive dataset was assembled. Extensive patient data collection facilitated the development of a sophisticated efficacy prediction model using deep learning algorithms.

Methods

Between 2018.01- 2022.11, data from 369 esophageal squamous cell carcinoma patients undergoing concurrent chemoradiotherapy at 2 medical centers were retrospectively analyzed. Dosimetric factors, including average and maximum doses for PTV, were derived from DVH. Clinical features such as gender, age, smoking and alcohol history, and clinical staging were also extracted. A total of 1349 radiomic features were extracted, and cluster analysis was utilized to identify radiomic features correlated with treatment efficacy. The deep learning algorithm was subjected to end-to-end training, amalgamating multiple data sources to establish a complex predictive model.

Results

In cluster analysis, a total of 112 highly correlated radiomic features were identified for further investigation. The model exhibited promising performance metrics on the validation set, with an accuracy of 0.802, sensitivity of 0.801 (95% CI [0.245, 0.987]), and specificity of 0.818 (95% CI [0.359, 0.995]). During five-fold cross-validation, the AUC ranged between 0.83 and 0.93, indicating robustness across multiple iterations of model evaluation.

Conclusions

This study introduces a comprehensive approach, leveraging deep learning techniques to integrate diverse data sources for predicting concurrent chemoradiotherapy efficacy in esophageal squamous cell carcinoma. The model not only demonstrates exceptional predictive accuracy but also pioneers the exploration of 3D dose distribution significance, offering novel insights for tailored treatment strategies.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Has not received any funding.

Disclosure

All authors have declared no conflicts of interest.