PREVIS: Predictive visual analytics of anatomical variability for radiotherapy decision support. (June 2021)

• Record Type: Journal Article
• Authors: Furmanová, Katarína; Muren, Ludvig P.; Casares-Magaz, Oscar; Moiseenko, Vitali; Einck, John P.; Pilskog, Sara; Raidou, Renata G.
• Is Part Of: Computers & graphics. Volume 97(2021)
• Page Start: 126
• Page End: 138
• Publication Date: 2021-06
• Subjects: Medical Visualization -- Visual Analytics -- Comparative Visualization -- Ensemble Visualization -- Radiotherapy Planning -- Cohort Study; Computer graphics -- Periodicals; 006.6
• DOI: 10.1016/j.cag.2021.04.010
• Languages: English
• ISSNs: 0097-8493

Abstract

Highlights: An application for exploration and prediction of pelvic organ variability in radiotherapy. Retrospective patient data are employed to predict anatomical changes in new incoming patients. The prediction is linked to treatment plan evaluation, supporting the selection of the optimal treatment. Usage scenarios and an evaluation; both conducted with eight domain experts. Graphical abstract: Abstract: Radiotherapy (RT) requires meticulous planning prior to treatment, where the RT plan is optimized with organ delineations on a pre-treatment Computed Tomography (CT) scan of the patient. The conventionally fractionated treatment usually lasts several weeks. Random changes (e.g., rectal and bladder filling in prostate cancer patients) and systematic changes (e.g., weight loss) occur while the patient is being treated. Therefore, the delivered dose distribution may deviate from the planned. Modern technology, in particular image guidance, allows to minimize these deviations, but risks for the patient remain. We present PREVIS : a visual analytics tool for (i) the exploration and prediction of changes in patient anatomy during the upcoming treatment, and (ii) the assessment of treatment strategies, with respect to the anticipated changes. Records of during-treatment changes from a retrospective imaging cohort with complete data are employed in PREVIS, to infer expected anatomical changes of new incoming patients with incomplete data, using a generative model. Abstracted representations of the retrospective cohort partitioning provide insight into an underlying automated clustering, showing main modes of variation for past patients. Interactive similarity representations support an informed selection of matching between new incoming patients and past patients. A Principal Component Analysis (PCA)-based generative model describes the predicted spatial probability distributions of the incoming patient's organs in the upcoming weeks of treatment, based on observations of past patients. The generative model is interactively linked to treatment plan evaluation, supporting the selection of the optimal treatment strategy. We present a usage scenario, demonstrating the applicability of PREVIS in a clinical research setting, and we evaluate our visual analytics tool with eight clinical researchers.