DeepCT-SI: Deep Learning Algorithm for Computed Tomography Source Identification

Large multi-center publicly available computed tomography (CT) datasets are important for artificial intelligence (AI) research in medicine. However, one CT scanner may create images that are of inadequate quality and identifying images from a given CT scanner can be a significant challenge. We propose a deep learning-based algorithm for CT Source Identification (DeepCT-SI). DeepCT-SI utilizes a four-layered convolutional neural network and a three-layered fully connected neural network. Our study involves a cohort of 2107 patients (mean age: 61.82, standard deviation(SD): 16.64) representing both sexes (male: 49%, female: 51%). We randomly partitioned the dataset into training/validation (n=1303, 80%) and test (n=361, 20%) datasets based on patients with no leakage of patient data between data partitions. We utilize 18 different physical CT scan machines, including four distinct manufacturers and eight distinct models. This diverse set of machines represented a broad range of imaging technology commonly employed in clinical practice. DeepCT-SI demonstrates superior prediction performance on the training dataset, achieving an accuracy of 99.75%, AUC of 99.17%, a precision of 98.49%, specificity of 98.88%, and sensitivity of 98.47%. DeepCT-SI evaluated on the test dataset, achieving an accuracy of 97.37%, AUC of 90.79%, precision of 83.16%, specificity of 98.47%, and sensitivity of 83.10%.