Knowledge-Based Three-Dimensional Dose Prediction for Tandem-And-Ovoid Brachytherapy
Purpose: To develop a knowledge-based voxel-wise dose prediction system using a convolution neural network for high-dose-rate brachytherapy cervical cancer treatments with a tandem-and-ovoid (T&O) applicator. Methods: A 3D U-NET was utilized to output dose predictions using organ-at-risk (OAR), high-risk clinical target volume (HRCTV), and possible source locations. A sample of previous T&O treatments comprising 397 cases (273 training:62 validation:62 test), HRCTV and OARs (bladder/rectum/sigmoid) was used. Structures and dose were interpolated to 1x1x2.5mm3 dose planes with two input channels (source positions, voxel identification) and one output channel for dose. We evaluated dose difference (ΔD)(xyz)=D_(actual)(x,y,z)-D_(predicted)(x,y,z) and dice similarity coefficients in all cohorts across the clinically-relevant dose range (20-130% of prescription), mean and standard deviation. We also examined discrete DVH metrics used for T&O plan quality assessment: HRCTV D_90%(dose to hottest 90% volume) and OAR D_2cc, with ΔD_x=D_(x,actual)-D_(x,predicted). Pearson correlation coefficient, standard deviation, and mean quantified model performance on the clinical metrics. Results: Voxel-wise dose difference accuracy for 20-130% dose range for training(test) ranges for mean (ΔD) and standard deviation for all voxels was [-0.3%+/-2.0% to +1.0%+/-12.0%] ([-0.1%+/-4% to +4.0%+/-26%]). Voxel-wise dice similarity coefficients for 20-130% dose ranged from [0.96, 0.91]([0.94, 0.87]). DVH metric prediction in the training (test) set were HRCTV(ΔD_90)=-0.19+/-0.55 Gy (-0.09+/-0.67 Gy), bladder(ΔD_2cc)=-0.06+/-0.54 Gy (-0.17+/-0.67 Gy), rectum(ΔD)_2cc=-0.03+/-0.36 Gy (-0.04+/-0.46 Gy), and sigmoid(ΔD_2cc)=-0.01+/-0.34 Gy (0.00+/-0.44 Gy). Conclusion: 3D knowledge-based dose predictions for T&O brachytherapy provide accurate voxel-level and DVH metric estimates.