KEY POINTS
- The retrospective single-institution dataset contained 619 lymphoma plans from 571 patients, spanning highly heterogeneous target locations and volumes. The original clinical VMAT plans were converted into standardized 15-field coplanar 6-MV sliding-window IMRT reference plans, with 61 plans reserved for independent testing.
- Two deep-learning approaches were compared: a conventional dose-only model and a multi-objective model combining dose supervision, fluence supervision, and dose–fluence–dose cycle consistency. Predicted fluence maps were imported directly into Eclipse, leaf-sequenced, and recalculated with AcurosXB without additional optimization or manual editing.
- Deliverability-aware training substantially improved target coverage. Mean PTV D98 increased from 78.59% with dose-only training to 90.10%, approaching 91.62% for the TPS reference plans. The median paired D98 improvement was +12.31 percentage points (Holm-corrected p<0.0001).
- Hotspot control also improved: mean PTV D2 decreased from 109.65% to 106.28%, while Dmax decreased from 113.27% to 110.32%. Median paired improvements were −3.43 percentage points for D2 and −3.18 percentage points for Dmax, both p<0.0001.
- The deliverability-aware model still did not fully match the TPS reference. Compared with the reference plans, median D98 remained 1.66 percentage points lower (p=0.014), while D2 and Dmax remained 1.60 and 4.01 percentage points higher, respectively.
- Using exploratory thresholds of D98 >90% and Dmax ≤110%, only 4.9% of dose-only plans satisfied both criteria versus 31.1% with the multi-objective model and 72.1% of TPS reference plans. Fluence-map generation remained fast at <1 second per map.
- The ablation analysis showed that fluence supervision was the main stabilizing component: adding it alone increased mean D98 from 78.59% to 88.27%. Cycle consistency without fluence supervision actually worsened performance, whereas combining both produced the best results. Generalizability remains uncertain because the reference plans used a fixed 15-field geometry and were standardized research replans rather than individually optimized clinical plans.
CLINICAL TAKEAWAY
For AI-based treatment planning, good dose prediction is not enough if the resulting modulation cannot survive sequencing and TPS recalculation. This work moves automated planning closer to a clinically meaningful endpoint by optimizing for deliverability, but plan quality still remained inferior to the TPS reference and external validation across other techniques and institutions is needed.