931P - Deep learning models for predicting short-term efficacy in locally advanced nasopharyngeal carcinoma

Background

Currently, there is a paucity of validated deep learning methodologies for predicting short-term efficacy in locally advanced nasopharyngeal carcinoma (NPC), both domestically and internationally. Furthermore, advanced techniques such as 2.5D feature extraction and Transformer model establishment remain underexplored in this context.

Methods

MRI images from locally advanced NPC patients diagnosed at two centers between July 2020 and June 2023 were retrospectively collected. Three state-of-the-art (SOTA) models (SegResNet, Unet, and UnetR models) were employed for automatic segmentation of all regions of interest (ROIs), thus generating an automatic segmentation model. Leveraging 2.5D deep learning techniques, tumor image features were extracted and integrated into a transfer learning framework. Subsequently, a Transformer-based model was designed, integrating all individual patient features. Three models were established: a clinical model, a Transformer model, and a fusion model. The predictive performance of these models was evaluated using receiver operating characteristic (ROC) curves, calibration curves, and decision curve analysis (DCA).

Results

Among the three models, the Transformer model demonstrated superior predictive performance in both the testing (AUC = 0.957) and validation cohorts (AUC = 0.893) compared to the clinical model. Moreover, the fusion model generally matched or surpassed the performance of the Transformer model, particularly in the testing cohort with an AUC of 0.874, indicating that the integration of clinical data with advanced machine learning models can enhance predictive capability.

Conclusions

The fusion model, which combines machine learning with clinical factors, exhibits promising efficacy in predicting short-term outcomes in locally advanced NPC patients following synchronous chemoradiotherapy.

Clinical trial identification

Editorial acknowledgement

Funding

Has not received any funding.

Disclosure

All authors have declared no conflicts of interest.