Generating intermediate slices with U-nets in craniofacial CT images

Abstract

AbstractAimThe Computer Tomography (CT) imaging equipment varies across facilities, leading to inconsistent image conditions. This poses challenges for deep learning analysis using collected CT images. To standardize the shape of the matrix, the creation of intermediate slice images with the same width is necessary. This study aimed to generate inter-slice images from two existing CT images.Materials and MethodsThe study utilized CT images from the Japanese Facial Bone Fracture CT Collection Project. The pixel values were converted to Hounsfield numbers and normalized. Three re-slice systems utilizing U-nets were developed: 1/3, 1/4, and 1/5. The datasets were divided into training and validation sets, and data augmentation techniques were applied. The U-net models were trained for 200 epochs. Validation was conducted using validation datasets. The generated images were compared to the corresponding original images using peak signal-to-noise ratio (PSNR), structural similarity (SSIM), and mean squared error (MSE) calculations. Results: Statistical analysis revealed significant differences between linear interpolation and U-net prediction in all indexes.ConclusionThe developed re-slice systems with U-net models showed practical value for making intermediate slice images from the existing images in the craniofacial area.