Predicting 2-Year Overall Survival in NSCLC from CT Scans Using 2D CNNs and Soft Attention

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate overall survival (OS) prediction in non-small cell lung cancer (NSCLC) is vital but hindered by complex 3D CT data and time-to-event outcomes.
• Conventional 3D Convolutional Neural Networks (CNNs) are computationally intensive and prone to overfitting, especially with limited annotated datasets.

Purpose of the Study:

• To develop a computationally efficient and accurate framework for NSCLC OS prediction.
• To overcome the limitations of 3D CNNs by leveraging 2D CT slice information with soft attention.

Main Methods:

• A lightweight framework aggregating 2D CT slice embeddings using soft attention to create a 3D patient representation.
• Feature extraction using EfficientNetB0 and time-to-event survival modeling with DeepHit.
• Validation on the LUNG1 dataset (415 patients) and transfer learning on the CLARO dataset.

Main Results:

• The proposed method demonstrated superior performance over 3D ResNet and other aggregation strategies, evidenced by a higher time-dependent concordance index (Ctd-index).
• Transfer learning significantly improved prediction accuracy on a smaller private dataset (CLARO).
• Achieved a substantially lower computational cost (54.3 GFLOPs) compared to ResNet3D-18 (2924.6 GFLOPs).

Conclusions:

• 2D CNNs combined with soft attention offer an effective and computationally efficient alternative to 3D CNNs for NSCLC OS prediction.
• The framework provides a practical solution for accurate survival prediction using readily available 2D CT data.
• This approach holds promise for improving clinical decision-making in NSCLC management.