COCONET: A Coordinate-Convolutional patch-based ResUnet for MR Pediatric Image Synthesis
Tongyao Wang, Yasheng Chen, Paul K. Commean, Cihat Eldeniz, Corinne M. Merrill, Gary B. Skolnick, Kamlesh Patel, Hongyu An · Proceedings on CD-ROM - International Society for Magnetic Resonance in Medicine. Scientific Meeting and Exhibition/Proceedings of the International Society for Magnetic Resonance in Medicine, Scientific Meeting and Exhibition · 2025
Motivation: CT is widely used for detecting pediatric cranial abnormalities but can increase the risk of cancer due to ionizing radiation. MR-synthesized pseudo-CT (pCT) is a safe alternative but challenging for infant data. The generalization of pCT to more than one magnetic field strength is needed for broad clinical adoption. Goal(s): We aim to develop a method to generate pCTs for children (0-18 years old) using MRI acquired at 1.5T and 3T. Approach: We proposed a 3D patch-based coordinate-convolutional ResUNet (COCONET) and utilized transfer learning to refine models for infants and different magnetic fields. Results: Our method produced pCTs similar to the gold-standard CTs. Impact: This study provides high-resolution pCT images from pediatric MRI imaging. It provides an alternative pediatric cranial bone imaging method free from ionizing radiation.