Room: Karl Dean Ballroom C
Purpose: The major limiting step in automating treatment plans is determining what to treat. Currently, radiation oncologists manually define beam apertures when treating with a conventional 4-field box technique. The purpose of this study is compare two state-of-the-art deep learning architectures to automatically define beam apertures for cervical cancer patients treated with a 4-field box technique.
Methods: Computed tomography scans from 310 cervical cancer patients were used in this study. Digitally-reconstructed radiographs (DRRs) were created for anterior-posterior, posterior-anterior, right-lateral, and left-lateral views. The DRR images, along with their corresponding physician-approved beam apertures, were used as inputs and ground-truth, respectively. Patients were randomly split into training(80%) and test(20%) sets, and individual models were trained for each DRR view. To investigate the optimal architecture parameters (# of layers, input image size, cost function, penalization parameters, drop-out, etc.) we ran a 3-fold cross-validation on our training dataset. Two deep learning architectures were investigated. First, we evaluated the performance of the U-net architecture, which allows for end-to-end training of the dataset. Second, we implemented a transfer learning segmentation approach where the first few layers used weights from the VGG19 architecture and then used up-convolutional layers to provide the predicted segmentation. Overlap and distance metrics between the predictions and ground-truth were calculated to evaluate each architectureâ€™s performance.
Results: The two architectures provided accurate segmentations in comparison to the physician-approved ground-truths. Average Dice similarity coefficients (DSC) were 0.97Â±0.02 and 0.95Â±0.02 and average mean surface distances were 2.7Â±2.5 mm and 4.2Â±1.7 mm for the U-net and VGG19 architectures, respectively.
Conclusion: The definition of beam apertures can be automated with high-accuracy using two independent deep learning approaches. While the predicted beam apertures in this work reflect patterns in our own clinical practice, this approach can be repeated with other datasets from other institutions to replicate their clinical practice.
Funding Support, Disclosures, and Conflict of Interest: This work was partially supported by NIH grant #UH2 CA20665