A simple and interpretable method to predict head and neck offline adaptation needs based on anatomical changes

Modern radiotherapy techniques do not automatically account for anatomical changes during treatment, which can lead to reduced target coverage, hotspots, or increased dose to healthy tissues. Adaptive Radiotherapy (ART) addresses this by updating the treatment plan based on the patient's current anatomy. In routine clinical practice, the decision to replan is often triggered by a suboptimal fit of the immobilization mask, identified by the radiation therapy technologist. Because this process is operator dependent and lacks methodological rigor, a quantitative decision-support method could substantially improve the replanning decision. Daily image-guided radiotherapy (IGRT) images provide valuable clinical information but often lack sufficient quality for accurate target contouring or reliable dose calculation to support straightforward clinical decision-making. This limitation motivates the need for a simpler, more practical, and scalable alternative.

This study develops an automated workflow to predict the need for replanning at every treatment fraction in head and neck (H&N) cancer patients using robust tools, thus requiring minimal manual validation while aiding the radio-oncologists in their decision-making.

A retrospective dataset of 21 H&N cancer patients treated on a Tomotherapy Radixact system (Accuray, USA) with kVCT daily imaging was analyzed, of whom 14 underwent replanning during treatment. For each patient, the planning CT (pCT) and daily images were processed in RayStation treatment planning system (RaySearch Laboratories AB, Sweden) via scripts. Rigid registrations were used to align daily images to the pCT, and a patient-specific region of interest (ROI) was defined to consistently capture the same anatomical region across all treatment fractions. ROI volumes were intersected with the external contour and converted to relative volumetric changes with respect to the pCT. By grouping patients according to whether a replanning occurred during treatment, differences in mean volumetric change between replanned and non-replanned patients were assessed at each treatment fraction. To automatically detect clinically relevant anatomical changes, a simple, interpretable model based on three volumetric features (current volume loss, short-term trend, and long-term trend) was evaluated using cross-validation. Thresholds defining the trigger condition were optimized using a tailored loss function that prioritizes early detection and penalizes false negatives more strongly than false positives. Model variants using different feature subsets and hyperparameters were compared, and the final configuration was selected prioritizing recall, followed by specificity and early flagging.

Significant differences in volumetric change between replanned and non-replanned patients emerged at fraction 6 and remained significant for most subsequent fractions. The smallest adjusted $p$ -values occurred around fractions 22-25. Using leave-one-out cross-validation, the best-performing model flagged patients to be replanned with 93% recall, doing so on average 8.4 fractions before clinical intervention.

Consistent differences in local tissue volume loss between replanned and non-replanned patients enabled the development of a simple automated workflow that triggers replanning for offline ART when learned volumetric thresholds are exceeded. This approach provides an objective complement to physician judgement and integrates seamlessly into existing IGRT workflows.