Room: Davidson Ballroom A
Purpose: Radiotherapy for head-and-neck(H&N) cancer is complex and time-consuming, where the planner spends days meticulously tuning the optimizer, cycling through iterations of potential plans without realizing when the best potential plan has been reached. Inexperienced centers may also not have a â€œgold standardâ€? reference to ensure the final product is on par with the standard-of-care. An accurate dose prediction model may solve these problems by providing high-quality templates against which a candidate plan may be compared.
Methods: We combined two state-of-the-art deep learning architectures, U-net and DenseNet, into a Hierarchically-Dense(HDense) U-net to capture local and global features(U-net) while having more efficient feature propagation(DenseNet). We collected plans from 120 H&N patients treated with VMAT, dividing the data into 80 training, 20 validation, and 20 testing patients. Using the PTVs and 22 OARs as input, we tuned and trained the deep network to predict the clinical dose, and compared our prediction against the standard U-net.
Results: On average, the deep learning model predicted the PTV coverage and max dose within 0.1%(D95), 1.6%(D98), 2.7%(D99), and 4.7%(Dmax) of the prescription dose. The organ-at-risk(OAR) dose was predicted within 7.4%(Dmax) and 5.9%(Dmean).The average PTV_homogeneity((D2-D98)/D50) and conformity ((VPTVâˆ©V100%Isodose)Â²/(VPTVÃ—V100%Isodose)) were predicted to be 0.09 and 0.76 (ground truth is 0.06 and 0.76). Overall, standard U-net is inferior to HDense, predicting within 0.1%(PTV_D95), 2.3%(PTV_D98), 3.2%(PTV_D99), 8.9%(PTV_Dmax), 9.36%(OAR_Dmax), 7.35%(OAR_Dmean) of the prescription dose, 0.15(PTV_homogeneity), and 0.73(conformity). Using the Wilcoxon-signed-rank-test, the improved prediction of HDense over standard U-net for OAR max/mean dose, PTV max dose, and homogeneity, is statistically significant with pâ‰¤0.01, but is insignificant for PTV coverage(D95,D98,D99)(p=0.11), and conformity(p=0.50).
Conclusion: We introduce a novel deep learning framework for accurately predicting clinical doses from patient contours. This can serve as clinical guidance for dosimetrists to achieve clinically acceptable plans with less physician feedback, saving time and raising the plan quality baseline.
Funding Support, Disclosures, and Conflict of Interest: This study was supported by the Cancer Prevention & Research Institute of Texas (CPRIT) IIRA RP150485.