2024 World Conference on Lung Cancer (WCLC) - Posters-P4.07G.02 Prediction of Relapse-free Survival of NSCLC Patients Through Multimodal Data Fusion Using Deep Learning Model

P4.07G.02 Prediction of Relapse-free Survival of NSCLC Patients Through Multimodal Data Fusion Using Deep Learning Model

Back to course

Pdf Summary

The study explores the prediction of two-year relapse-free survival in non-small cell lung cancer (NSCLC) patients by integrating clinical data and CT images using a deep learning model named IC-TCA-Net. Traditional models often rely on single data modalities, offering limited perspectives. By combining multimodal data, this approach seeks to enhance prediction accuracy.<br /><br />The research involved 166 NSCLC patients who underwent surgical resection, with data comprising 17 clinical variables and preoperative chest CT scans. The IC-TCA-Net model integrates deep image features, extracted through a tumor-centric attention module, and clinical features for recurrence prediction using a classifier with fully connected layers. Through univariate and multivariate logistic regression analyses, three key clinical features were selected: N stage, vascular invasion, and lactate dehydrogenase (LDH) levels.<br /><br />The study employed 3-fold cross-validation for dataset configuration, using the Adam optimizer with a batch size of 16 and a learning rate of 10e-4. The dataset was split randomly with stratified sampling to ensure balanced representation across training, validation, and testing. <br /><br />The IC-TCA-Net demonstrated improved performance compared to models using only CT images. Its accuracy was recorded at 80.17%, with balanced accuracy at 80.87%, sensitivity at 83.86%, specificity at 77.89%, and an AUC of 0.86. This indicates a superior capacity to predict relapse-free survival compared to a model limited to CT image features alone.<br /><br />The conclusion highlights that integrating CT and clinical data within a deep learning framework significantly improves the model's ability to predict relapse-free survival, suggesting that these data types offer complementary insights, which the deep learning model efficiently weighs to enhance prediction accuracy. This approach could potentially lead to better-informed clinical decisions and tailored patient management strategies.

Asset Subtitle

Kyongmin Beck

Meta Tag

Speaker Kyongmin Beck

Topic Early-Stage NSCLC

Keywords

NSCLC

relapse-free survival

IC-TCA-Net

deep learning

multimodal data

clinical data

CT images

tumor-centric attention

prediction accuracy

patient management