Advancing Normal Tissue Complication Probability Modeling with Supervised Contrastive Learning for Predicting...

Area of Science:

• Radiation oncology
• Medical physics
• Machine learning in healthcare

Background:

• Normal tissue complication probability (NTCP) modeling using dose-volume histograms (DVHs) faces challenges like high dimensionality and multicollinearity.
• Classical classification methods are limited by overlapping DVH profiles in patients with different toxicity outcomes.
• Accurate prediction of radiation-induced toxicity is crucial for personalized cancer treatment planning.

Purpose of the Study:

• To introduce SC-NTCP, a supervised contrastive learning framework for transforming DVH data into a compact, separable latent representation.
• To optimize DVH data representation for predicting osteoradionecrosis (ORN) in head and neck cancer patients.
• To improve the accuracy and interpretability of NTCP modeling compared to traditional approaches.

Main Methods:

• Developed SC-NTCP, a supervised contrastive learning framework to create a latent representation of DVH data.
• Maximized intra-class similarity and inter-class separability within the embedding space.
• Benchmarked SC-NTCP against logistic regression, SVM, MLP, and CNN using a cohort of head and neck cancer patients, incorporating clinical covariates.

Main Results:

• SC-NTCP achieved superior discrimination for ORN prediction with an AUC of 0.77.
• The framework demonstrated improved calibration and enhanced interpretability through gradient-based feature attribution.
• Integration of clinical covariates further augmented the predictive performance of SC-NTCP.

Conclusions:

• SC-NTCP offers a principled and interpretable approach to robust radiation toxicity prediction, overcoming limitations of raw DVH data.
• The method enhances NTCP modeling for osteoradionecrosis in head and neck cancer.
• SC-NTCP has the potential to inform personalized treatment planning and improve clinical outcomes in radiation oncology.