Standardized 3D Medical Image Translation Framework Evaluates Seven Generative Models
A new approach has been developed for translating 3D medical images, allowing for virtual scans without needing additional images. This research looks at seven different generative models, which include three GANs: Pix2Pix, CycleGAN, and SRGAN, as well as four latent generative models: Latent Diffusion Model, Latent Diffusion Model+ControlNet, Brownian Bridge, and Flow Matching. These models are tested on eleven datasets covering the head/neck, lung, and pelvis areas. The framework maintains uniformity in preprocessing, data splitting, inference, and evaluation for cancer imaging. Unlike past methods that usually dealt with 2D images and lacked clinical testing, this study aims to address those shortcomings through a detailed comparison.
Key facts
- Framework standardizes 3D medical I2I translation evaluation
- Seven generative models compared: Pix2Pix, CycleGAN, SRGAN, Latent Diffusion Model, Latent Diffusion Model+ControlNet, Brownian Bridge, Flow Matching
- Eleven datasets across three anatomical regions: head/neck, lung, pelvis
- Focus on oncological imaging
- Previous methods mostly 2D, isolated tasks, no clinical validation
Entities
—