End-to-end deep learning pipeline for automated IMRT planning in lymphoma: Feasibility and limitations

Background: Intensity-modulated radiation therapy (IMRT) is widely used for lymphoma because a multileaf collimator (MLC) can sculpt dose around irregular targets. Planning, however, is time-consuming and difficult to standardize. Fluence map prediction (FMP) using deep learning (DL) could potentially be used to automate planning, but it has not been evaluated in full-body lymphoma, where prescriptions, organ-at-risk (OAR) sets, and anatomical presentations vary substantially. Here, we evaluate multiple DL configurations to determine whether they can be adapted to lymphoma and generate IMRT plans that are deliverable in a clinical treatment planning system (TPS). Purpose: To investigate the feasibility and limitations of DL–based dose and fluence prediction methods in the heterogeneous setting of lymphoma. Methods: A single-institution lymphoma cohort (571 patients, 619 plans) was replanned from clinical volumetric modulated arc therapy (VMAT) to a consistent 15-field IMRT reference using the Eclipse Scripting API (ESAPI), yielding 9285 beam-level pairs. We implemented a two-stage DL pipeline that first predicts beam-wise dose from CT and planning target volume (PTV), then maps dose to fluence, and finally computes the corresponding dose in the TPS after leaf sequencing. We evaluated several architectural and training variants, from replication-focused models to configurations with a clinical loss based on dose–volume histogram (DVH) aimed at moving beyond simple plan replication. For each configuration, we analysed how these choices influenced predicted dose realism and the quality of the resulting plans. Results: In Stage 1 (dose prediction), replication-focused DL models were numerically inferior to the IMRT and VMAT references, while DVH-driven variants yielded superior target DVHs. After fluence mapping and TPS sequencing, all configurations produced machine-deliverable IMRT plans but with lower target coverage and higher hot spots than the references; fluence was largely preserved by MLC sequencing, indicating that current limitations mainly stem from the predicted dose distributions and the learned dose-to-fluence mapping. Conclusions: DL–based dose and fluence prediction can generate fast, machine-deliverable IMRT plans for heterogeneous lymphoma that lie in the clinical ballpark, but they remain dosimetrically inferior to manual IMRT and VMAT, highlighting the current limitations of these methods for automated planning in this setting.