Deep learning signatures reveal multiscale intratumor heterogeneity associated with biological functions and survival in recurrent nasopharyngeal carcinoma.

ABSTRACT

PURPOSE: How to discriminate different risks of recurrent nasopharyngeal carcinoma (rNPC) patients and guide individual treatment has become of great importance. This study aimed to explore the associations between deep learning signatures and biological functions as well as survival in (rNPC) patients. METHODS: A total of 420 rNPC patients with PET/CT imaging and follow-up of overall survival (OS) were retrospectively enrolled. All patients were randomly divided into a training set (n = 269) and test set (n = 151) with a 64 ratio. We constructed multi-modality deep learning signatures from PET and CT images with a light-weighted deep convolutional neural network EfficienetNet-lite0 and survival loss DeepSurvLoss. An integrated nomogram was constructed incorporating clinical factors and deep learning signatures from PET/CT. Clinical nomogram and single-modality deep learning nomograms were also built for comparison. Furthermore, the association between biological functions and survival risks generated from an integrated nomogram was analyzed by RNA sequencing (RNA-seq). RESULTS: The C-index of the integrated nomogram incorporating age, rT-stage, and deep learning PET/CT signature was 0.741 (95% CI 0.688-0.794) in the training set and 0.732 (95% CI 0.679-0.785) in the test set. The nomogram stratified patients into two groups with high risk and low risk in both the training set and test set with hazard ratios (HR) of 4.56 (95% CI 2.80-7.42, p < 0.001) and 4.05 (95% CI 2.21-7.43, p < 0.001), respectively. The C-index of the integrated nomogram was significantly higher than the clinical nomogram and single-modality nomograms. When stratified by sex, N-stage, or EBV DNA, risk prediction of our integrated nomogram was valid in all patient subgroups. Further subgroup analysis showed that patients with a low-risk could benefit from surgery and re-irradiation, while there was no difference in survival rates between patients treated by chemotherapy in the high-risk and low-risk groups. RNA sequencing (RNA-seq) of data further explored the mechanism of high- and low-risk patients from the genetic and molecular level. CONCLUSION: Our study demonstrated that PET/CT-based deep learning signatures showed satisfactory prognostic predictive performance in rNPC patients. The nomogram incorporating deep learning signatures successfully divided patients into different risks and had great potential to guide individual treatment patients with a low-risk were supposed to be treated with surgery and re-irradiation, while for high-risk patients, the application of palliative chemotherapy may be sufficient.