Time-Driven Survival Analysis from FDG-PET/CT in Non-Small Cell Lung Cancer.

Area of Science:

• Radiomics and Medical Imaging
• Computational Oncology
• Deep Learning in Healthcare

Background:

• Accurate prediction of overall survival (OS) is crucial for personalized treatment planning in Non-Small Cell Lung Cancer (NSCLC).
• Current prognostic models often lack the precision needed for individualized patient care.
• Automated prediction using medical imaging offers a promising avenue for enhanced prognostics.

Purpose of the Study:

• To develop and evaluate a deep regression framework for predicting OS in NSCLC patients.
• To integrate tissue-wise FDG-PET/CT projections and temporal data for improved survival prediction.
• To compare the proposed framework against a baseline method utilizing only imaging data.

Main Methods:

• A ResNet-50 backbone was used to generate image embeddings from FDG-PET/CT projections.
• Temporal data (time horizon in days) was combined with image embeddings for OS probability prediction.
• The framework was developed on the U-CAN cohort (n=556) and validated on a test set (n=292).

Main Results:

• The proposed model incorporating temporal data improved OS prediction accuracy (AUC +4.3%) over the baseline.
• An ensemble of imaging and clinical data achieved the highest performance (0.788).
• The model enabled effective patient risk stratification and highlighted tumor regions via saliency analysis.

Conclusions:

• The developed method provides an automated framework for time-dependent OS prediction in NSCLC.
• Combining imaging and tabular data significantly enhances survival prediction accuracy.
• This approach holds potential for improving patient prognostics and guiding personalized treatment strategies.