Deep-RSIv2: An Efficient Content-Free Deep Learning Approach for Radiographs' Manufacturer and Model Identification

Artificial Intelligence (AI)-driven digital world is vulnerable against AI-based threats, particularly deepfakes, that have made image forgery to become ubiquitous. Digital alteration, or AI-aided generation of radiology images adversely influence patient care. We collected 3875 patients(
$\mathrm{n}=4125$
radiographs), including chest (1509 patients,
$\mathrm{n}=3353$
radiographs, mean age=64.99,female:63%) and thoracic spine (2366 patients,
$\mathrm{n}=8644$
radiographs, mean age=67, female: 33%). Data were randomly divided by patient into training/validation(80%), test(20%). We develop Deep-RSIv2, an efficient content-free deep-learning framework using light three convolutional layers followed by fully connected neural network layers to authenticate radiographs by manufacturers with AUC=[chest(3 classes): 95.85%, thoracic spine(3 classes): 93.87%]; and models with AUC=[chest(5 classes): 93.87%, thoracic spine(4 classes): 91.36%].