Predicting lung cancer survival with attention-based CT slices combination

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Accurate prognosis for Non-Small Cell Lung Cancer (NSCLC) is vital for effective patient management and treatment planning.
• Despite deep learning advancements, overall survival (OS) prediction in NSCLC remains a challenge.
• Current methods may not fully exploit the potential of deep learning for survival analysis using medical imaging.

Purpose of the Study:

• To develop and validate a novel deep learning methodology for predicting 2-year OS in NSCLC patients from CT scans.
• To integrate CT scan features with a soft attention mechanism to identify prognostically relevant image slices.
• To compare the proposed method against benchmark 3D networks and assess its adaptability with different backbones and transfer learning.

Main Methods:

• A novel deep learning approach combining EfficientNetB0 for CT slice representation and a soft attention mechanism to focus on relevant slices for risk assessment.
• Validation using the public LUNG1 dataset and a private dataset, with comparisons against benchmark 3D convolutional neural networks.
• Exploration of transfer learning on the private dataset to evaluate performance in data-limited settings.

Main Results:

• The proposed method achieved a mean C-index of 0.584 on the LUNG1 dataset, outperforming benchmark 3D networks.
• Combining 2D slice representations to form a 3D volume representation proved more effective for OS prediction than traditional 3D approaches.
• Transfer learning significantly improved performance on the private dataset, increasing the C-index by 0.076.

Conclusions:

• The developed deep learning methodology effectively predicts 2-year overall survival in NSCLC patients using CT scans.
• The attention-based approach for integrating 2D slice information into a 3D representation is superior to traditional 3D methods for this task.
• Transfer learning offers a viable strategy to enhance prediction accuracy in scenarios with limited patient data.