AI-assisted partial breast planning generates acceptable plans in six of seven end-to-end cases

AI-generated partial breast plans met all constraints in six of seven end-to-end cases, but fixed beam geometry caused frequent heart-dose violations.

KEY POINTS

  • This retrospective single-institution feasibility study from the University of Pennsylvania evaluated an AI-assisted planning workflow for accelerated partial breast irradiation (APBI) delivered to 30 Gy in five fractions. Three separate, nonoverlapping cohorts were assessed: 33 patients for automated target segmentation, 40 for knowledge-based VMAT autoplanning, and seven for a fully integrated segmentation-to-plan pipeline. Automated plans were compared with clinically generated manual optimization plans using institutional target and OAR criteria.
  • The segmentation component used an nnU-Net trained on 533 previously treated APBI patients, following target definitions aligned with the Florence trial. From 48 consecutive patients available for testing, 33 with clearly visualizable lumpectomy cavities were included. The lumpectomy cavity was designated as GTV, with the PTV derived using a 2-cm expansion cropped 5 mm from the skin and according to institutional anatomical boundaries.
  • AI-generated contours achieved mean GTV Dice similarity of 0.78 ± 0.09 and PTV Dice similarity of 0.83 ± 0.07 compared with physician contours. 97% of GTVs achieved Dice >0.5, while 67% of PTVs exceeded Dice 0.8. The higher PTV agreement partly reflected expansion and anatomical cropping rather than superior identification of the original cavity. Important failure modes included incorrect laterality associated with surgical clips, over-contouring of clips, and difficulty distinguishing seroma from dense breast tissue.
  • In the 40-patient autoplanning cohort, generated plans achieved PTV D95 of 31.51 ± 0.18 Gy versus 31.16 ± 0.20 Gy with manual optimization (p<0.001). OAR sparing improved for ipsilateral lung V5Gy (13.03% vs 18.77%; p=0.010) and contralateral breast D0.03cc (1.46 vs 2.40 Gy; p=0.001). Automated plan generation required 9.8 ± 1.8 minutes, demonstrating the speed of knowledge-based optimization under standardized planning conditions.
  • Fixed arc geometry was the principal limitation of standalone autoplanning. Only 24/40 automated plans (60%) met all institutional DVH constraints, compared with 37/40 manually optimized plans (92.5%). Fifteen automated plans exceeded the heart V3Gy ≤10% constraint, and one failed the hotspot criterion. Mean heart V3Gy was 12.39% with autoplanning versus 1.74% manually (p=0.014). Among the 24 acceptable automated plans, heart V3Gy was comparable between approaches (1.09% vs 0.69%; p=0.245).
  • In the seven-patient end-to-end cohort, six automated plans met all institutional DVH requirements, versus seven manually generated plans. PTV D95 was 29.20 ± 0.40 Gy with automation versus 29.77 ± 0.12 Gy manually (p=0.063). The single unsuccessful automated plan exceeded the ipsilateral breast V28.5Gy constraint (33% vs the 30% limit). Manual plans achieved lower heart mean dose (0.54 vs 0.73 Gy; p=0.016) and ipsilateral lung V5Gy (11.44% vs 16.03%; p=0.031), while automated plans reduced contralateral breast D0.03cc (1.43 vs 1.98 Gy; p=0.016).
  • The integrated workflow reduced the reported interval from simulation to plan acceptance to less than one hour, compared with a historical average of approximately nine days for conventional clinical planning. However, the automated measurement primarily reflected processing and plan-generation steps, whereas the manual interval included scheduling, waiting time, review, and revisions, so these figures are not equivalent measures of hands-on time savings. The study did not evaluate clinical toxicity, actual treatment delivery, or prospective deployment, and the authors recommend physician review of both contours and final DVHs before treatment approval.

CLINICAL TAKEAWAY

Integrating automated lumpectomy cavity segmentation with knowledge-based planning could substantially reduce the technical workload and turnaround time for five-fraction APBI, with most tested end-to-end plans meeting institutional requirements. However, frequent cardiac dose violations with fixed beam geometry and errors in cavity recognition demonstrate that clinically reliable automation still requires patient-specific optimization logic, independent validation, and mandatory physician and physics oversight.

SOURCE

Advances in Radiation Oncology

Browse more research Suggest a correction