2018/12/21 by Aaron Babier, Babier, Aaron, Rafid Mahmood +7
Medicine · Physics and Astronomy · #Advanced Radiotherapy Techniques #FOS: Physical sciences #Lung Cancer Diagnosis and Treatment #Medical Physics (physics.med-ph) #Radiomics and Machine Learning in Medical Imaging
paper · pdf · doi:10.48550/arxiv.1812.09309
openalex publication_date 2018/12/21 · openalex created_date 2022/08/01 · openalex updated_date 2026/07/28
We develop a knowledge-based automated planning (KBAP) pipeline that\ngenerates treatment plans using deep neural network architectures for\npredicting 3D doses. Our pipeline consisted of a generative adversarial network\n(GAN) to predict dose from a CT image followed by two optimization models to\nlearn objective function weights and generate fluence-based plans,\nrespectively. We investigated three different GAN models. The first two models\npredicted dose for each axial slice independently. One predicted dose as a RGB\ncolor map, while the other predicted a scalar value for dose directly. The\nthird GAN model predicted scalar doses for the full 3D CT image at once,\nconsidering correlations between adjacent CT slices. For all models, we also\ninvestigated the impact of scaling the GAN predictions before optimization.\nEach GAN model was trained on 130 previously delivered oropharyngeal treatment\nplans. Performance was tested on 87 out-of-sample plans, by evaluating using\nclinical planning criteria and compared to their corresponding clinical plans.\n The best performing KBAP plans were generated with the 3D GAN, which\npredicted dose values followed by scaling. These plans satisfied close to 77%\nof all clinical criteria, compared to the clinical plans, which satisfied 64%\nof all criteria. Additionally, these KBAP plans satisfied the same criteria as\nthe clinical plans 84% more frequently compared to the 2D GAN model using RGB\ndose prediction. The 3D GAN predictions were also more similar to the final\nplan compared to the other approaches, as it better captured the vertical\ndosimetric relationship between adjacent axial slices. The deliverable plans\nbetter captured implicit constraints associated with physical deliverability.\nOverall, our final plans had superior performance in satisfying clinical\ncriteria and generated more realistic predictions compared to the previous\nstate-of-the-art.\n