DINO-Reg

Official repository for the MICCAI 2024 paper:
DINO-Reg: General Purpose Image Encoder for Training-Free Multi-modal Deformable Medical Image Registration

---

DINO-Reg is a simple, powerful, and training-free framework for deformable medical image registration.
Built on top of the DINOv2 ViT backbone, it enables multi-modal registration with general-purpose image features.

---

📁 Repository Structure

• inference_l2rmrct.py — Main script for inference.
• models/ — First run will download DINOv2 model here (dinov2_vitl14_reg4_pretrain.pth).
• dinov2/ — Contains config files and model architecture.
• sample_dataset_dir/ — Placeholder directory for input images (user needs to populate this).

---

⚙️ Environment Setup

We recommend using the same environment as DINOv2.
Follow the DINOv2 setup instructions to install the necessary dependencies.

---

🔧 How to Run

1. Populate your dataset
   Add your input image pairs into the sample_dataset_dir/ directory. Modify the csv files accordingly.

   pairs_Tr.csv should contain the filenames, and structures.csv contains the index for the organ segmentations that need to be evaluated.

2. Run Inference

   python inference_l2rmrct.py