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 (Delta D)(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 Delta D_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 (Delta D) 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(Delta D_90)=-0.19+/-0.55 Gy (-0.09+/-0.67 Gy), bladder(Delta D_2cc)=-0.06+/-0.54 Gy (-0.17+/-0.67 Gy), rectum(Delta D)_2cc=-0.03+/-0.36 Gy (-0.04+/-0.46 Gy), and sigmoid(Delta D_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.
Download