Cone-beam computed tomography (CBCT) is an imaging exam used in endodontics to diagnose diseases and evaluate internal root anatomy, however, one of its limitations is the presence ofartefacts that hamper the diagnostic efficacy. Although the literature presents different techniquesto reduce artefacts, till this moment, there are no effective methods that contribute to its reduction. Deep convolutional neural network (DCNN) has shown promising approaches to reduce thepresence of artefacts arising from high-density materials, however, besides the scientific literatureonly presents preliminary data, with neural networks with low training by few CBCT images,there is no information about its impact on diagnostic accuracy of clinical tasks such as rootfracture and perforation. Therefore, this study aims to create a DCNN to reduce artefacts arisingfrom high-density materials and low X-ray dose, based in CBCT scans, evaluate the impact in thediagnostic accuracy of different clinical tasks and analyze and compare its efficacy with othersartefact reduction methods. The Ad-hoc DCNN initially design will be corrected in thecomputational facilities of Mésocentre Moulon (Ruche - Lenovo supercomputer ofCentraleSupélec and ENS Paris Saclay) considering its preliminary results and all CBCT imagesof the imaging phantom will be reinserted in the DCNN for network validation and training. TheCBCT images will be corrected by the trained DCNN and will be analyzed subjectively by fourobservers to indicate the presence of vertical root fracture and root perforation with a 5-pointscale. And, additionally, the images will be analyzed objectively by one observer, that will obtainmeasurements of the signal-to-noise ratio and gray value homogeneity of a homogeneous area.All the images will be evaluated in the CBCT software's OsiriX MD v.7.5.1 and e-Vol DX andwill be posteriorly compared.
News published in Agência FAPESP Newsletter about the scholarship: