Pix2Rep-v2: Data-Efficient Representation Learning for Dense Medical Imaging Applications
Paper • 2609.01427 • Published
This modelcard contains the weights of Pix2Rep-v2 pre-training on cine MRI data (combined ACDC+MnMs+MnMs2 training sets) as in Pix2Rep-v2: Data-Efficient Representation Learning for Dense Medical Imaging Applications. The two backbones (Pix2Rep-v2+U-Net and Pix2Rep-v2+SwinUNETR) are released.
Note: SwinUNETR backbone has been trained with GroupNorm for the projection head (you need to specify it from the command line pretraining.backbone.params.norm=group when instanciating the model), while all other models use BatchNorm for the projection head.
Please take a look at paper https://arxiv.org/abs/2609.01427 and official GitHub repository https://github.com/BioMedTP/pix2rep-v2 for more information