핵심 개념
Deep learning models improve survival prediction by learning ordered representations based on survival times.
초록
1. Abstract
- Predicting survival crucial for cancer patients.
- Deep learning models struggle with regression-aware feature representations.
- Proposed SurvRNC method orders representations based on survival times.
2. Introduction
- Survival prediction vital in medical care.
- Complex task due to various factors and censored data.
- CoxPH model widely used but has limitations.
3. Methodology
- Develop deep neural network for survival prediction using multimodal data.
- Feature encoder and survival predictor components explained.
- Introduce Surv Rank-N-Contrast loss function for ordered feature representation.
4. Experimental Setup
- Experiment with DeepMTLR and DeepHit models.
- Compare performance with other methods on HECKTOR dataset.
- Standardized testing environment ensures impartiality.
5. Results and Discussion
- Including LSurvRNC boosts performance in all models.
- Proposed method outperforms state-of-the-art techniques on HECKTOR dataset.
- Effectiveness of SurvRNC demonstrated through ablation study and test set results.
6. Conclusion
- SurvRNC improves survival prediction by ordering latent representations based on time-to-event labels.