Data-Efficient and Explainable Multimodal Survival Prediction in NSCLC Using Deep Image Embeddings, Clinical...

Area of Science:

• Oncology
• Radiology
• Data Science

Background:

• Non-small cell lung cancer (NSCLC) survival prediction is challenging, especially with limited data.
• Deep learning models can struggle with generalization in small datasets.
• A multimodal and explainable approach is needed to integrate imaging and clinical data.

Purpose of the Study:

• To develop a data-efficient, multimodal, and explainable framework for NSCLC survival prediction.
• To integrate computed tomography (CT)-derived imaging information with clinical variables.
• To improve the generalization of survival prediction models in limited-sample settings.

Main Methods:

• Utilized the NSCLC Radiomics (LUNG1) dataset with CT images, segmentations, and clinical data.
• Extracted deep imaging embeddings using pretrained RadImageNet-InceptionV3 and fused them with engineered features and clinical variables.
• Employed principal component analysis for feature compression and trained gradient-boosted tree models (CatBoost, XGBoost, LightGBM) with SHAP for interpretability.

Main Results:

• The CatBoost model achieved a C-index of 0.655 for three-class survival stratification.
• The LightGBM model achieved a C-index of 0.576 for continuous survival regression.
• Clinical variables were the primary prognostic indicators, with deep image embeddings offering complementary information, particularly for distinguishing short- and long-term survival.

Conclusions:

• The proposed framework demonstrates a feasible, data-efficient, and explainable image-to-tabular approach for NSCLC survival prediction.
• Combining pretrained CT embeddings, clinical data, gradient-boosted trees, and SHAP analysis shows promise for limited-sample survival modeling.
• External validation is crucial before clinical translation of this approach.