DIP-UP

deepMRI collection

DIP-UP provides a DIP-based enhancement mechanism for pre-trained 3D CNN models in MRI phase unwrapping. It refines existing network outputs using two unsupervised constraints: Laplacian loss and total variation (TV) loss — no additional labeled data needed.

Pipeline: (1) Train a base phase unwrapping network → (2) Apply DIP enhancement at test time.

16–24 GB GPU memory recommended. Tested on Windows 10.

---

Overview

Pipeline

---

Requirements

• Python 3.7+, PyTorch
• NVIDIA GPU (16–24 GB VRAM recommended)

---

Checkpoints and Test Data

• Pre-trained models (PHU-NET3D and PhaseNet3D): Dropbox

• Test data (simulation: 10 ms TE, σ=0.1; in vivo: 5.8 ms TE): Dropbox

---

Usage

Step 1 – Prepare data

Two network variants are provided:

Network	Data loader	Input channels
PHU-NET3D	TrainingDataLoad_ResidueLoss_2Chan.py	image_file, lap_file, Label_file
PhaseNet3D	TrainingDataLoad_ResidueLoss_1Chan.py	image_file, Label_file

Step 2 – Train base network

Edit the following variables in the training script:

Variable	Description
DATA_DIRECTORY	Root directory of training dataset
DATA_LIST_PATH	Index file (e.g. test_IDs_28800.txt)
ModelFolder	Path to save trained model checkpoints
ModelName	Name for saved model files

Step 3 – DIP enhancement (test time)

Edit and run the inference script:

Variable	Description
SourceDir	Directory of inference data
ReconType	Simulation or InVivo
SaveDir_NIFTI	Output directory for .nii results

Learning rate options:

• Variable: decays every 10 epochs during DIP
• Constant: uses default fixed value

---

⬆ top  |  deepMRI collection