Session Item

Recent studies have shown that deep learning (DL) is promising for distant metastasis (DM) and overall survival (OS) prognosis in H&N cancer with segmented PET or CT. However, the predictive power could be diminished by the variation in primary and lymph node gross tumor volume (GTV) segmentation. Moreover, the potential of joint modality prognosis remains to be investigated. This study aimed to explore prognosis without GTV segmentation, to extend the single modality prognosis to joint PET/CT, and to investigate the predictive performance with different modality inputs.

We implemented 3D-Resnet and extended it to time-to-event analysis by using an existing survival model to incorporate all censoring and survival information. Publicly available CTs and PETs from 4 different Canadian hospitals (293 patients) and MAASTRO clinic (74 patients) were used for training by 3-fold cross-validation (CV). For independent testing, we used 110 patients from a collaborating institution. All the 477 patients received radiotherapy (RT) or chemo-RT as primary treatment. Different modality inputs were trained and tested, including PET or CT with (by masking the images) or without primary and lymph node GTV contours. For the joint PET/CT prognosis, we set each modality as a separate input channel. The predictive performances were evaluated by Harrell’s Concordance Index (HCI) and Kaplan-Meier curves.

Table 1 displays the best CV HCIs and their corresponding testing HCIs of different models. Compared with previously published works, there were improvements in CV and testing HCIs for both DM and OS. Generally, PET achieved better predictive performance than CT in both CV and testing cohorts. Besides, PET-only outperformed PET with GTV contour (PET-GTV), indicating that GTV segmentation might not be indispensable in PET-based prognosis. Comparatively, GTV contours could be important for CT-based prognosis, since the CT-only testing HCIs were poor. Although higher than the CT-only model, the HCIs of joint PET/CT were not improved compared to the PET-only model. This could be explained by the redundancy over different imaging modalities and might also suggest that multi-channel input is not optimal to exploit multi-modality information. Fig. 1 shows the Kaplan-Meier curves of the PET-only and CT-GTV models (the best choice for each modality), demonstrating the significant stratification capability of the trained models for the testing cohort. 

DL-based DM and OS time-to-event models showed predictive capability and could benefit personalized RT treatment. The high accuracy of the PET-only model also suggested GTV segmentation might be less relevant for PET-based prognosis. The predictive performance of joint PET/CT could not be directly improved from the PET-only model and remains to be explored in future studies.