Large-scale dose evaluation of deep learning organ contours in head-and-neck radiotherapy by leveraging existing plans

Graphical abstract Workflow for automated plan optimization and use-case of evaluating the effect of automated contours on dose. By reusing original plan (POG) parameters, we made a plan for both the manual contours (PMC) and automated contours(PAC), shown with yellow and blue colors respectively. Dashed lines indicate the evaluation workflow where both doses were evaluated on the manual contours. Pink, maroon and orange contours are used to represent the manual, automated and PTV (DL1) contours respectively. Finally, we used manual contours to compute dose metrics and normal tissue complication probability (NTCP) models and compare all plans.


Supplementary Material A. Data Acquisition
The CT scans of our dataset had a dimension of 512 x 512 pixels in the spatial plane with a pixel spacing in the range of [0.92-1.36, 0.92-1.36]mm.Each CT slice was 2mm thick and each scan had between [128,199] slices.The scans were acquired from a Brilliance Big Bore (Philips Healthcare, Ohio, USA) with 120kV and 250mAs.Post acquisition, 64% of patients had Orthopedic Metal Artifact Reduction (O-MAR) processing done.

Supplementary Material B. Automated Contours
The auto-contouring model of RayStation 10B (results in Table B.1 and Table B.2) first performed registration of the chosen CT scan using an atlas of CTs to narrow down CT size so it fits within the graphical processing unit (GPU) used for deep learning.Once registered, the mid-point of each OAR is detected and a 3D bounding box is cropped around that.This cropped area is then passed to a neural net trained for contouring that specific OAR.Each OAR-specific neural net is based on the UNet segmentation architecture whose output is a 3D probabilistic mask for that OAR.As a post-processing step, smoothing is performed on the surfaces of OARs.

Supplementary Material C. Automated Planning
For automated planning, we replicated the beam setup, OAR/target objectives for both photon and proton as per our institutions clinical head-and-neck protocol.
For photon (Table C.3), our VMAT plans are made on an isotropic dose grid of 0.2cm The photon beams were commissioned on an Elekta Synergy system with Agility multi-leaf collimator.
For proton (Table C.4), our IMPT plans are made on an isotropic dose grid of 0.3cm.This dose is delivered using pencil beam scanning (PBS) on a Varian ProBeam machine.
Step  C.3: Our 4-step emulation of the manual photon optimization process of our clinic.In each step, we also optimize for the objectives of the previous steps.We use VDT as an abbreviation for the phrase "value determined by treatment planner".The → indicates that the weight is modified at the end of Step 4..Here DL1/DL2 stands for electives/boost regions of the tumor and prescription refers to a value of cGy that was assigned to a region-of-interest (RoI).Here "Other Organs" refers to Cochlea (L/R), Parotid (L/R).Submandibular (L/R), Muscle Constrictor (S/M/I), Cricopharyngeus, Larynx (SG), Glottic Area, Trachea, Esophagus and Oral Cavity.The rows shown here are created as objectives in our clinic's treatment planning solution. Step Figure F.7: This figure shows CT scans of photon (a-f) and proton (g-j) patients overlayed with a dose distribution as well as PTV (DL1) (orange), PTV (DL2) (blue), manual (pink) and automated (maroon) contours.Each example shows the P OG , P M C and P AC plans from left to right.The dose metric in the sub-captions compares the absolute percentage difference of P M C − P AC .
The model was trained using Tensorflow, an open-source deep neural net software package.During training, rotations, translations and elastic deformations were used to augment the training data.Details on patient cohort were not made public by the manufacturer.

Table
TableC.4:Our 4-step emulation of the manual proton optimization process of our clinic.In each step, we also optimize for the objectives of the previous steps.We use VDT as an abbreviation for the phrase "value determined by treatment planner".The → indicates that the weight is modified at the end of Step 4..Here DL1/DL2 stands for elective/boost regions of the CTV and prescription refers to a value in cGy that was assigned to a region-of-interest (RoI)."Organ Set 1" refers to Mandible, Brainstem, Spinal Cord, Esophagus, Trachea, Larynx (SG), Trachea and Glottic Area, while "Organ Set 2" refers to Parotid (L/R), Submandibular (L/R), Muscle Constrictor (S/M/I), and Oral Cavity.The * mark is used to indicate those objectives which are robustly optimized.The rows shown here are created as objectives in our clinic's treatment planning solution.