Casting wider than U-Nets: evaluating the role of novel machine learning applications in radiation therapy

Radiation oncology has proven an early adopter of machine learning (ML) and deep learning (DL) owing to its image-centered workflow and data richness. DL auto-segmentation of normal structures is already guideline-endorsed, with self-configuring nnU-Net delivering near-state-of-the-art accuracy and widely adopted. In this review, we explore advances in ML algorithms in processing and interpreting imaging-based data beyond normal tissue auto-segmentation with applications upstream and downstream of this task, highlighting training data volume and best-in-class architectures. Applications discussed range from diagnostics with "virtual biopsies" and automated staging to radiotherapy planning with gross tumor volume and clinical target volume (CTV) prediction. We illustrate the developments of cutting-edge transformer-based designs and foundation models, but also the important role of lightweight, task-specific classical ML models. Data scarcity may be eased by developments in few-shot learning (FSL) and transfer learning strategies. We also consider how end-to-end automated CTV delineation may require an agentic approach to mimic the reasoning of a radiation oncologist, capable of processing multi-modal data with clinical context. Together, these advancements point toward a future where ML will become increasingly central to radiation oncology practice. However, implementation of these tools beyond normal tissue auto-segmentation remains limited and will require careful assessment and validation in the clinic.