Unpaired

.gitignore

@@ -1,4 +1,14 @@
 *.pyc
 *pycache*
 saved_info/*
-saved_info
+saved_info
+slurm/outputs/*
+synthetic_data/*
+.DS_Store
+pretrained/*
+wandb/*
+data/*
+results/*
+debug.ipynb
+.env
+old/*

README.md

@@ -1,4 +1,4 @@
-# SynDiff
+# SynDiff (Original ReadMe)
 
 Official PyTorch implementation of SynDiff described in the [paper](https://ieeexplore.ieee.org/document/10167641).
 

create_datasets.py

@@ -0,0 +1,151 @@
+import sys
+path_to_pip_installs = "/tmp/test_env"
+if path_to_pip_installs not in sys.path:
+    sys.path.insert(0, path_to_pip_installs)
+
+import matplotlib.pyplot as plt
+import numpy as np
+import pandas as pd
+import h5py
+import torch
+from torch.utils.data import DataLoader
+from pytorch_dataset import HDF5ContrastDataset
+
+def save_dataset_as_mat(dataset, out_file, var='data_fs', compression='gzip'):
+    """Save so that LoadDataSet() shows the image upright."""
+
+    imgs = []
+    for i in range(len(dataset)):
+        img = dataset[i]['image']          # (1,256,256)  channel‑first
+        if isinstance(img, torch.Tensor):
+            img = img.cpu().numpy()
+        img = img.squeeze(0) .T            # <-- transpose here (W, H)
+        imgs.append(img.astype(np.float32))
+
+    imgs = np.stack(imgs, axis=0)          # (N,256,256)
+    with h5py.File(out_file, 'w') as f:
+        f.create_dataset(var, data=imgs,
+                         dtype='float32', compression=compression)
+
+    print(f'✔ Saved {len(imgs)} slices to {out_file} (will load as (N,1,256,256))')
+
+def get_patient_z_dim_combinations(overview_df_filtered, contrast_list):
+    contrast_groups = overview_df_filtered[overview_df_filtered["contrast"].isin(contrast_list)].groupby('contrast').apply(
+        lambda g: set(zip(g['patient_id'], g['z_dim']))
+    )
+
+    # Step 2: Take the intersection of all sets
+    common_combinations = set.intersection(*contrast_groups)
+
+    # Step 3: Convert back to a DataFrame if needed
+    result_df = pd.DataFrame(list(common_combinations), columns=['patient_id', 'z_dim'])
+
+    return result_df
+
+def load_dataset(base_path, contrast, image_type, cfg_transform, split="train", image_dim=0):
+    # Load the synthetic dataset
+    cfg_filters = {
+        "contrast__in": [contrast],
+        "non_zero": True,
+        "image_dim": 0,
+        "image_type": image_type,
+        "split":split,
+    }
+    dataset = HDF5ContrastDataset(
+        hdf5_path=f"{base_path}/data_{contrast}.h5",
+        filter=cfg_filters,
+        transform=cfg_transform,
+        stage="eval",  
+    )
+    return dataset
+
+def save_both_dataset_as_mat(dataset1, dataset2, out_file1, out_file2, var='data_fs', compression='gzip'):
+    """Save so that LoadDataSet() shows the image upright."""
+
+    imgs_1 = []
+    imgs_2 = []
+    already_vistied_j = set()  # To avoid duplicate processing of dataset2
+    for i in range(len(dataset1)):
+        img_1 = dataset1[i]['image']          # (1,256,256)  channel‑first
+        img_1_metadata = dataset1[i]['metadata']
+        for j in range(len(dataset2)):
+            if j in already_vistied_j:
+                continue
+            img_2_metadata = dataset2[j]['metadata']
+            if img_1_metadata['patient_id'] == img_2_metadata['patient_id'] and img_1_metadata['z_dim'] == img_2_metadata['z_dim']:
+                img_2 = dataset2[j]['image']
+                if isinstance(img_1, torch.Tensor):
+                    img_1 = img_1.cpu().numpy()
+                img_1 = img_1.squeeze(0).T  # <-- transpose here (W, H)
+                imgs_1.append(img_1.astype(np.float32))
+                if isinstance(img_2, torch.Tensor):
+                    img_2 = img_2.cpu().numpy()
+                img_2 = img_2.squeeze(0).T # <-- transpose here
+                imgs_2.append(img_2.astype(np.float32))
+                already_vistied_j.add(j)
+                break
+
+    imgs_1 = np.stack(imgs_1, axis=0) # (N,256,256)
+    with h5py.File(out_file1, 'w') as f:
+        f.create_dataset(var, data=imgs_1,
+                         dtype='float32', compression=compression)
+
+    print(f'Saved {len(imgs_1)} slices to {out_file1} (will load as (N,1,256,256))')
+
+    imgs_2 = np.stack(imgs_2, axis=0) # (N,256,256)
+    with h5py.File(out_file2, 'w') as f:
+        f.create_dataset(var, data=imgs_2,
+                         dtype='float32', compression=compression)
+
+    print(f'Saved {len(imgs_2)} slices to {out_file2} (will load as (N,1,256,256))')
+
+
+def create_datasets(contrast1, contrast2, image_type1, image_type2, cfg_transform, base_path, output_path):
+
+    dataset1_train = load_dataset(base_path, contrast1, image_type1, cfg_transform, split="train")
+    dataset2_train = load_dataset(base_path, contrast2, image_type2, cfg_transform, split="train")
+
+    dataset1_val = load_dataset(base_path, contrast1, image_type1, cfg_transform, split="val")
+    dataset2_val = load_dataset(base_path, contrast2, image_type2, cfg_transform, split="val")
+
+    dataset1_test = load_dataset(base_path, contrast1, image_type1, cfg_transform, split="test")
+    dataset2_test = load_dataset(base_path, contrast2, image_type2, cfg_transform, split="test")
+
+    out_file_1_train = f"{output_path}/{contrast1}_{contrast2}_train.h5"
+    out_file_2_train = f"{output_path}/{contrast2}_train.h5"
+
+    out_file_1_val = f"{output_path}/{contrast1}_{contrast2}_val.h5"
+    out_file_2_val = f"{output_path}/{contrast2}_val.h5"
+
+    out_file_1_test = f"{output_path}/{contrast1}_{contrast2}_test.h5"
+    out_file_2_test = f"{output_path}/{contrast2}_test.h5"
+
+    save_both_dataset_as_mat(dataset1_train, dataset2_train, out_file_1_train, out_file_2_train, var='data_fs', compression='gzip')
+    save_both_dataset_as_mat(dataset1_val, dataset2_val, out_file_1_val, out_file_2_val, var='data_fs', compression='gzip')
+    save_both_dataset_as_mat(dataset1_test, dataset2_test, out_file_1_test, out_file_2_test, var='data_fs', compression='gzip')
+
+
+def main():
+    cfg_transform = {
+    "eval": [
+        {
+        "GroupMinMaxNormalize": {
+            "stats_path": "/home/students/studweilc1/SynthRegGAN/data/minmax_values.json"
+        }
+        },
+
+    ]
+    }
+
+    base_path = "/home/students/studweilc1/SynthRegGAN/data"
+    output_path = "/home/students/studweilc1/SynDiff/data/my_data_group"
+
+    contrast1 = "DIXON"
+    contrast2 = "BOLD"
+    image_type1 = "W"
+    image_type2 = "s"
+
+    create_datasets(contrast1, contrast2, image_type1, image_type2, cfg_transform, base_path, output_path)
+
+if __name__ == "__main__":
+    main()

dataset.py

@@ -5,10 +5,10 @@
 
 def CreateDatasetSynthesis(phase, input_path, contrast1 = 'T1', contrast2 = 'T2'):
 
-    target_file = input_path + "/data_{}_{}.mat".format(phase, contrast1)
+    target_file = input_path + "/data_{}_{}.h5".format(phase, contrast1)
     data_fs_s1=LoadDataSet(target_file)
 
-    target_file = input_path + "/data_{}_{}.mat".format(phase, contrast2)
+    target_file = input_path + "/data_{}_{}.h5".format(phase, contrast2)
     data_fs_s2=LoadDataSet(target_file)
 
     dataset=torch.utils.data.TensorDataset(torch.from_numpy(data_fs_s1),torch.from_numpy(data_fs_s2))