Strategic Automation for Gynecological Brachytherapy Planning: Optimizer Enhanced Deep Learning and Distance-Based Techniques
Abstract
Purpose
To automate the cervical cancer brachytherapy (CC-BT) planning process using two
Methods
(1) deep learning (DL) and (2) distance-based (DB) initialization approaches, both enhanced with an autograd optimizer (DL+OPTAG and DB+OPTAG, respectively), for generating human-like personalized plans.
Methods
Method 1 was trained/validated on 100 treated MR-based high-dose-rate CC-BT plans (following EMBRACE II protocol) to predict dwell times, with anatomy and dwell positions as inputs to learn the treated dose distribution. Method 2 initialized dwell times based on source distance to relevant structures. Dwell times generated by both methods were fed into an Adam optimizer with a nonlinear continuous loss function on structural dose-volume histogram (DVH) constraints. The developed DL+OPTAG and DB+OPTAG methods were tested on an additional 30 treated plans (ground truth, GT; using Geneva and Venezia applicators). Evaluations employed DVH metrics (critical organ D2cc, CTV-HR D90), dose conformity index (COIN), total reference air-kerma (TRAK), applicator weighting, and physician review.
Results
GT CTV-HR volumes were 10-108 cc, receiving D90 of 5-8.5 Gy/fraction (planning aim of 7.75 Gy/fraction; 100% intracavitary/interstitial cases). Both methods improved the GT plans, with an increase in structures meeting ideal and acceptable dosimetry. Compared to GT, both methods [DL+OPTAG, DB+OPTAG] yielded a mean dose (Gy) change/fraction for CTV-HR of [0.22±0.39, 0.25±0.43], bladder [-0.19±0.44, -0.21±0.45], rectum [-0.31±0.58, -0.35±0.61], sigmoid [-0.03±0.49, -0.05±0.50], and bowel [-0.06±0.38, -0.10±0.37]. Overall difference (%) in TRAK from GT was [-4, -5], with change in weights (%) for tandem being [+2, +2], vaginal [-11, -13], and needles [+9, +11]. Furthermore, conformity improved using both techniques; COIN [+7%, +9%].
Conclusion
The novel methods effectively replicated manual planning and generated clinically acceptable plans in <5min. As DB+OPTAG achieves comparable plans with reduced complexity, it could eliminate the need for a DL approach, which entails specific expertise and higher resource requirements.