A multi-kernel and multi-scale learning based deep ensemble model for predicting recurrence of non-small cell lung...

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Predicting non-small cell lung cancer (NSCLC) recurrence before treatment is crucial for personalized medicine.
• Deep learning, particularly convolutional neural networks (CNNs), shows promise in cancer informatics and time-to-event prediction.
• Existing CNN models often rely on single 2D CT images or 3D volumes, potentially missing valuable information.

Purpose of the Study:

• To develop and evaluate a deep learning-based ensemble network for predicting NSCLC recurrence.
• To leverage multi-scale inputs and feature fusion from multiple 2D CNN models to improve prediction performance.
• To identify high-risk NSCLC patients for potential adjuvant treatment strategies.

Main Methods:

• A dataset of 530 NSCLC patients was used.
• An ensemble network was designed, integrating multiple 2D CNN models with varying input slices, scales, and convolutional kernels (5x5 and 6x6).
• A deep learning-based feature fusion model was employed as the ensemble strategy.

Main Results:

• The proposed ensemble network achieved an accuracy of 69.62%, AUC of 72.5%, F1 score of 70.12%, and recall of 70.81%.
• The model demonstrated competitive performance compared to existing 2D and 3D-CNN models in benchmark studies.
• The ensemble approach effectively captured local and peritumoral features from multi-scale inputs.

Conclusions:

• The developed deep learning ensemble network shows potential for accurate NSCLC recurrence prediction.
• This method can provide more comprehensive information than single-slice or single-model approaches.
• The model serves as a potential tool for identifying NSCLC patients who may benefit from adjuvant therapy.