Multi-institutional generalizability of the CDM loss framework

Establish the generalizability of the clinical DVH metric loss (CDM loss)–based 3D head-and-neck radiotherapy dose prediction framework, trained on single-institution data, to datasets from other institutions and planning protocols by determining whether performance and satisfaction of clinical constraints are maintained across institutions.

Background

The study proposes a clinical DVH metric loss (CDM loss) that directly optimizes clinically used dose–volume metrics and introduces a bit-mask ROI encoding to improve computational efficiency. The method is evaluated on a retrospective cohort of 174 head-and-neck patients from a single institution.

Because all training and testing data come from one institution following a unified planning protocol, the extent to which the CDM loss–based framework generalizes to other institutions remains uncertain. Validating performance across institutions is necessary to confirm robustness and clinical applicability beyond the development site.