Ensemble of Convolutional Neural Networks for Sparse-View Cone-Beam Computed Tomography

Risks related to excessive exposure of patients to ionizing radiation are a significant concern in the medical community. Several approaches based on Convolutional Neural Networks (CNNs) have been proposed to develop safer and more reliable sparse-view Computed Tomography (SVCT) systems. Most of those solutions process tomographic data within 2D slices individually. However, recent works have shown that 3D models - that exploit data correlation among adjacent slices - can outperform previous 2D models. Once the kernel size in most of those 3D models is not bigger than 5 x 5 x 5, such inter-slice exploration is restricted to a limited neighborhood, resulting in minor inter-slice analysis during the training/validation phase. To efficiently exploit data correlation among the coronal, axial, and sagittal views of the SVCT volume, we propose an ensemble of four 2D CNNs. Three of them are used to process the orthogonal SVCT volume views separately, and the fourth CNN combines the outputs from the previous three networks. Since our final architecture is highly deep, we also present a training method in stages to avoid the non-convergence of the deepest layers. We conducted experiments using head cone-beam Computed Tomography (CBCT) scans extensively used in imaged guided radiotherapy (IGTR) during brain tumor treatment. Our method presented superior results in reducing reconstruction artifacts of SVCT volumes compared to the state-of-the-art 2D and 3D models.