Comparison of convolutional neural network training strategies for cone-beam CT image segmentation

• DOI: 10.1016/j.cmpb.2021.106192
• Publication Date: 2021-05-20T14:27:55Z
• Authors: Jordi Minnema; Jan Wolff; Juha Koivisto; Felix Lucka; Kees Joost Batenburg; Tymour Forouzanfar; Maureen van Eijna...
• Keywords: Similarity (geometry)

Over the past decade, convolutional neural networks (CNNs) have revolutionized field of medical image segmentation. Prompted by developments in computational resources and availability large datasets, a wide variety different two-dimensional (2D) three-dimensional (3D) CNN training strategies been proposed. However, systematic comparison impact these on segmentation performance is still lacking. Therefore, this study aimed to compare eight strategies, namely 2D (axial, sagittal coronal slices), 2.5D (3 5 adjacent majority voting, randomly oriented cross-sections 3D patches. These were used train U-Net an MS-D network for simulated cone-beam computed tomography (CBCT) images comprising randomly-placed non-overlapping cylinders experimental CBCT anthropomorphic phantom heads. The resulting performances quantitatively compared calculating Dice similarity coefficients. In addition, all segmented gold standard converted into virtual models using orientation-based surface comparisons. strategy that generally resulted best both was voting. When employing can be optimized slices are perpendicular predominant orientation anatomical structure interest. Such spatial features should taken account when choosing or developing novel results will help clinicians engineers choose most-suited