Predicting Overall Survival of NSCLC Patients with Clinical, Radiomics and Deep Learning Features.

Area of Science:

• Oncology
• Medical Imaging
• Radiotherapy

Background:

• Accurate overall survival (OS) estimation is crucial for non-small cell lung cancer (NSCLC) treatment planning.
• Previous studies highlighted the potential of radiomics and deep learning (DL) in improving survival prediction.

Purpose of the Study:

• To evaluate if a model integrating clinical, radiomics, DL, and dosimetric features outperforms models using only subsets of these data for NSCLC OS prediction.
• To assess the added value of multi-modal data integration in predicting patient outcomes after radiotherapy.

Main Methods:

• A cohort of 219 NSCLC patients' pre-treatment lung CT scans and clinical data were analyzed.
• Radiomics, DL (3D ResNet), and dosimetric features were extracted and combined with clinical data.
• An ensemble model (XGB and NN classifiers) was developed and evaluated using various feature combinations.

Main Results:

• The model using only clinical features achieved an AUC of 0.71 and 72.73% accuracy.
• The integrated model combining clinical, radiomics, dose, and DL features demonstrated superior performance with an AUC of 0.84 and 88.64% accuracy.
• World Health Organisation Performance Status was identified as the most important factor in the combined model.

Conclusions:

• Integrating radiomics, dosimetric, and DL features with clinical data significantly enhances the prediction accuracy of OS in NSCLC patients receiving radiotherapy.
• This enhanced predictive accuracy supports personalized, risk-based treatment planning, potentially leading to improved patient outcomes and quality of life.