Foundation model based prediction of lung cancer survival using temporal changes in dual time point CT scans.
Authors
Affiliations (7)
Affiliations (7)
- Department of Radiation Oncology, Dana-Farber Cancer Institute/Brigham and Women's Hospital, Harvard Medical School, Boston, MA, USA.
- Artificial Intelligence in Medicine (AIM) Program, Mass General Brigham, Harvard Medical School, Harvard Institutes of Medicine, Boston, USA.
- Radiology and Nuclear Medicine, CARIM and GROW, Maastricht University, Maastricht, the Netherlands.
- BC Cancer, Kelowna, British Columbia, Canada.
- University of British Columbia, Kelowna, British Columbia, Canada.
- Department of Radiation Oncology, Dana-Farber Cancer Institute/Brigham and Women's Hospital, Harvard Medical School, Boston, MA, USA. [email protected].
- Artificial Intelligence in Medicine (AIM) Program, Mass General Brigham, Harvard Medical School, Harvard Institutes of Medicine, Boston, USA. [email protected].
Abstract
Lung cancer remains a significant cause of mortality, with non-small cell lung cancer (NSCLC) representing most cases. Currently, clinical data based models fall short in predicting survival while more advanced deep learning based image models require vast amounts of data and are often limited to predictions based on single time points. This study uses dual time point CT scans and features derived from a foundation model to predict survival. A dataset containing 102 NSCLC patients treated with radiation therapy was used, with each patient having both pre-treatment and post-treatment CT scans. A foundation model applied to the scans generated high-dimensional feature vectors and these vectors were then further summarized. Statistical analyses, including random forest and gradient boosted survival models, were then used to predict survival. The results demonstrated that temporal changes in feature vectors, specifically the Euclidean distance and element-wise subtracted feature vectors, can offer improved prediction of survival over single-time point features and clinical data.