Foundation Features for Non-Small Cell Lung Cancer Recurrence Prediction
Abstract
Purpose
Foundation models are general-purpose, adaptable deep learning models trained on vast amounts of data. Despite unprecedented performance in language and vision tasks, their utility for predicting patient outcomes remains to be explored. This retrospective study investigates the utility of features extracted by foundation models for predicting two-year lung cancer recurrence type compared to conventional radiomic and clinical features.
Methods
357 newly diagnosed early-stage non-small cell lung cancer (NSCLC) patients treated with stereotactic ablative radiotherapy between 2006 and 2021 across five regional centres were eligible. A foundation model for cancer imaging biomarkers was used to extract features from tumour regions on pre-treatment computed tomography simulations. Imaging Biomarker EXplorer was used to extract radiomic features from gross tumour volumes. Patient charts were reviewed to identify clinical features and two-year recurrence type: no recurrence, locoregional recurrence, metastatic recurrence, or new primary. Features were selected using Spearman correlation coefficients, and visualized using t-SNE. Multi-class XGBoost models were trained using an 80%/20% split to predict recurrence type based on radiomic, foundation, or clinical features. Model performance was evaluated using balanced accuracy, weighted precision, weighted recall, and weighted F1-score. A dataset of 27 patients from a sixth regional centre were exclusively used for external validation.
Results
Spearman rank coefficient correlations were moderately strong between some foundation and radiomic features. No clear classification clusters were identified by t-SNE. Evaluating on the withheld test and external validation sets, models trained on foundation features alone had comparable performance to those trained on clinical features, but performance dropped when trained on both feature types. Model performance improved when trained on both radiomic and clinical features.
Conclusion
Foundation features were not found to be highly predictive of NSCLC recurrence, either alone or in combination with other features. Future work will evaluate the consistency of multi-institutional foundation features.