Speaker
Description
Artificial intelligence (AI) is increasingly applied in radiotherapy for tasks such as tumor segmentation and longitudinal response assessment. These methods, however, often assume complete high-quality imaging, which is not always available in clinical datasets. Missing or incomplete scans, particularly contrast-enhanced MRI, limit both inclusive AI model development and multi-institutional studies.
This talk presents recent works addressing these challenges. First, deep learning methods for radiotherapy longitudinal imaging analysis are described, highlighting performance in automated assessment of therapy outcomes under standard imaging conditions. Building on this foundation, the limitations caused by missing contrast-enhanced MRI are addressed through a contrast-aware generative modeling framework that synthesizes contrast-enhanced T1-weighted images from routinely acquired non-contrast sequences. The framework supports both single-timepoint and longitudinal imaging configurations, and is validated on multi-institutional datasets, demonstrating high fidelity and preservation of tumor enhancement patterns.
Finally, the efficacy of synthesized images for downstream tasks, including automated tumor detection and segmentation, is demonstrated, enabling reliable AI workflows even when key imaging modalities are unavailable. This work highlights a path toward data-robust AI systems in radiotherapy, bridging the gap between algorithmic performance and real-world applicability.