Purpose: Brachytherapy treatment planning for cervical cancer is performed real-time while patients are sedated and can take over an hour. There are few tools to automatically generate high-quality plans. While several groups have developed 3D knowledge-based dose predictions using neural networks, few have directly converted predictions into deliverable treatments. Building on prior work in 3D brachytherapy dose estimation, we developed an optimizer that would take brachytherapy dose distributions as input and output dwell times, to lay the foundation for automated, knowledge-based brachytherapy treatment planning.
Methods: A dose rate kernel was produced by exporting 3D dose for a single dwell position from the treatment planning system and normalizing by dwell time. By translating and rotating this kernel to each dwell position in an applicator, scaling by time for that position and summing over all dwell positions, dose for the full applicator was computed (Dcalc). Tandem-and-ovoid dwell positions, actual dwell times and 3D doses (Dref) were exported from a commercial planning system for ten previously-treated patients. To mimic Dref, dwell times were initialized to 15s and updated iteratively to minimize the mean squared error between Dcalc and Dref, computed using all voxels with Dref 80-150% of prescription dose. A COBYLA optimizer with a tolerance of 1e-3 was coded in Python. Final optimized dose (normalized to prescription) and dwell times were compared to actual values using mean absolute differences (MAD) over all voxels, dwell times, organ D2cc, CTV V100 and D90, and Dice coefficients for 100% isodose contours.
Results: Mean(range) of dose MAD over all voxels=1.5(0.7-3.4)%, dwell time MAD=3.58(1.98-8.27)s (0.76(0.34-1.79)% relative to total plan time), Dice=0.99(0.98-0.99). MAD in bladder/rectum/sigmoid D2cc= 0.5%/0.6%/0.4% and CTV D90/V100=0.64%/0.38%.
Conclusion: Our optimizer accurately reproduced dose distributions and dwell times of actual treatment plans, and could be used to generate automated plans from 3D dose predictions in the future.