Missing-Aware Multimodal Survival Prediction for NSCLC
Researchers have developed an innovative survival framework for Non-Small Cell Lung Cancer (NSCLC) that handles missing data by combining Computed Tomography (CT), Whole-Slide Histopathology Images (WSI), and clinical variables. This new approach uses Foundation Models (FMs) to pull out features from each type of data and includes a method that allows for combining these different data sources, even when some information is missing. The design ensures that every patient’s data is utilized during both training and evaluation, avoiding the common practice of excluding incomplete cases. This technique is particularly important for dealing with the challenges of small patient groups and missing data, which often limit the effectiveness of Multimodal Deep Learning (MDL) in predicting survival outcomes for unresectable stage II-III NSCLC.
Key facts
- Framework combines CT, WSI, and clinical variables
- Uses Foundation Models for feature extraction
- Missing-aware encoding enables fusion under incomplete modalities
- No patients dropped during training or inference
- Intermediate fusion outperforms conventional methods
- Targets unresectable stage II-III NSCLC
- Addresses small cohorts and missing modalities
- Published on arXiv as 2601.10386
Entities
Institutions
- arXiv