Predicting Non-Small-Cell Lung Cancer Survival after Curative Surgery via Deep Learning of Diffusion MRI
Jung Won Moon1, Ehwa Yang2, Jae-Hun Kim2
1Department of Radiology, Kangnam Sacred Heart Hospital, Hallym University School of Medicine, Seoul 07441, Republic of Korea.
Diagnostics (Basel, Switzerland)
|August 12, 2023
Summary
Deep learning models using diffusion-weighted imaging (DWI) and apparent diffusion coefficient (ADC) maps show promise for predicting non-small cell lung cancer survival. Combining DWI and ADC data significantly improves prediction accuracy over DWI alone.
Area of Science:
- Radiology and Medical Imaging
- Artificial Intelligence in Medicine
- Oncology
Background:
- Accurate survival prediction is crucial for non-small cell lung cancer (NSCLC) patient management.
- Diffusion-weighted imaging (DWI) offers insights into tissue microstructure, potentially aiding prognosis.
- Deep learning presents a novel approach to analyze complex imaging data for predictive modeling.
Purpose of the Study:
- To evaluate the predictive power of a deep learning survival model utilizing diffusion-weighted images (DWI) in NSCLC patients.
- To compare the performance of models using various combinations of DWI and apparent diffusion coefficient (ADC) parameters.
- To assess the added value of demographic features in the survival prediction model.
Main Methods:
- A deep learning survival model, based on a pre-trained VGG-16 network, was developed for 100 NSCLC patients.
- Input data included preoperative DWI (b-values 0, 100, 700 sec/mm²) and derived ADC maps (ADC₀-₁₀₀, ADC₁₀₀-₇₀₀, ADC₀-₁₀₀-₇₀₀), along with demographic features.
- Model performance was evaluated using 10-fold cross-validation, comparing accuracy and AUC for predicting 5-year survival.
Main Results:
- The best predictive performance was achieved using a combination of DWI₀, ADC₀-₁₀₀, and ADC₀-₁₀₀-₇₀₀, yielding 92% accuracy and an AUC of 0.904.
- Models incorporating ADC parameters significantly outperformed those using only DWI (p < 0.05).
- While demographic features improved models with DWI alone, their impact was less prominent when added to combined DWI and ADC models.
Conclusions:
- Deep learning holds potential for improving survival prediction in lung cancer patients.
- Integrating apparent diffusion coefficient (ADC) parameters derived from DWI enhances model performance compared to using DWI data solely.
- The study highlights the value of functional imaging parameters in deep learning for oncological prognostication.
