Updates

README.md

@@ -12,4 +12,88 @@ pretty_name: 3D Brain Structure MRI Autoencoder
 ---
 
 ## 🧠 Model Summary
-Coming soon.
+# brain2vec
+An autoencoder model for brain structure T1 MRIs based on [Brain Latent Progression](https://github.com/LemuelPuglisi/BrLP/tree/main). The autoencoder takes in a 3d MRI NIfTI file and compresses to 1200 latent dimensions before reconstructing the image. The loss functions for training the autoencoder are:
+- [L1Loss](https://pytorch.org/docs/stable/generated/torch.nn.L1Loss.html)
+- [KLDivergenceLoss](https://pytorch.org/docs/stable/generated/torch.nn.KLDivLoss.html)
+- [PatchAdversarialLoss](https://docs.monai.io/en/stable/losses.html#patchadversarialloss)
+- [PerceptualLoss](https://docs.monai.io/en/stable/losses.html#perceptualloss)
+
+# License
+MIT License
+
+Copyright (c) 2025
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+
+# Training data
+[Radiata brain-structure](https://huggingface.co/datasets/radiata-ai/brain-structure): 3066 scans from 2085 individuals in the 'train' split. Mean age = 45.1 +- 24.5, including 2847 scans from cognitively normal subjects and 219 scans from individuals with an Alzheimer's disease clinical diagnosis.
+
+# Example usage
+```
+# get brain2vec model repository
+git clone https://huggingface.co/radiata-ai/brain2vec
+cd brain2vec
+
+# set up virtual environemt
+python3 -m venv venv_brain2vec
+source venv_brain2vec/bin/activate
+
+# install Python libraries
+pip install -r requirements.txt
+
+# create the csv file listing the scan paths and other info
+python create_csv.py
+
+mkdir ae_cache
+mkdir ae_output
+
+# train the model
+nohup python brain2vec.py train \
+  --dataset_csv /home/ubuntu/brain2vec/inputs.csv \
+  --cache_dir   ./ae_cache \
+  --output_dir  ./ae_output \
+  --n_epochs    10 \
+> train_log.txt 2>&1 &
+```
+
+# Methods
+transform:
+(80, 96, 80)
+pixdim=2
+10 epochs
+    max_batch_size: int = 2,
+    batch_size: int = 16,
+    lr: float = 1e-4,
+
+# References
+Puglisi
+Pinaya
+
+# Citation
+```
+@dataset{Radiata-Brain-Structure,
+  author    = {Jesse Brown and Clayton Young},
+  title     = {Brain-Structure: Processed Structural MRI Brain Scans Across the Lifespan},
+  year      = {2025},
+  url          = {https://huggingface.co/datasets/radiata-ai/brain-structure},
+  note      = {Version 1.0},
+  publisher    = {Hugging Face}
+}
+```

autoencoder-ep-4.pth → autoencoder_final.pth

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d83a2a0ed04a16f4908e91c2c8aab3b20b4f9a763dd838600baba07e694c6b94
-size 55126081
+oid sha256:85c85a2a826d4667983e66ebb4f9ee6aad91cad1928bd05373dfcee0d8c1e18c
+size 55071236

brlp_lite.py → brain2vec.py

@@ -1,6 +1,6 @@
 # MIT License
 
-# Copyright (c) 2024
+# Copyright (c) 2025
 
 # Permission is hereby granted, free of charge, to any person obtaining a copy
 # of this software and associated documentation files (the "Software"), to deal
@@ -85,7 +85,7 @@ INPUT_SHAPE_1p5mm = (122, 146, 122)
 INPUT_SHAPE_AE = (80, 96, 80)
 
 # Latent shape of the autoencoder 
-LATENT_SHAPE_AE = (1, 10, 12, 10)   
+LATENT_SHAPE_AE = (1, 10, 12, 10)
 
 
 def load_if(checkpoints_path: Optional[str], network: nn.Module) -> nn.Module:
@@ -409,7 +409,7 @@ def train(
         batch_size=max_batch_size,
         shuffle=True,
         persistent_workers=True,
-        pin_memory=True
+        pin_memory=True,
     )
 
     print('Device is %s' %(device))

create_csv.py

@@ -0,0 +1,39 @@
+#!/usr/bin/env python3
+import os
+import pandas as pd
+from datasets import load_dataset
+
+def row_to_dict(row, split_name):
+    return {
+        "image_uid": row["id"],
+        "age": int(row["metadata"]["age"]),
+        "sex": 1 if row["metadata"]["sex"].lower() == "male" else 2,
+        "image_path": os.path.abspath(row["nii_filepath"]),
+        "split": split_name
+    }
+
+def main():
+    # Load the datasets
+    ds_train = load_dataset("radiata-ai/brain-structure", split="train", trust_remote_code=True)
+    ds_val = load_dataset("radiata-ai/brain-structure", split="validation", trust_remote_code=True)
+    ds_test = load_dataset("radiata-ai/brain-structure", split="test", trust_remote_code=True)
+
+    rows = []
+
+    # Process each split
+    for data_row in ds_train:
+        rows.append(row_to_dict(data_row, "train"))
+    for data_row in ds_val:
+        rows.append(row_to_dict(data_row, "validation"))
+    for data_row in ds_test:
+        rows.append(row_to_dict(data_row, "test"))
+
+    # Create a DataFrame and write it to CSV
+    df = pd.DataFrame(rows)
+    output_csv = "inputs.csv"
+    df.to_csv(output_csv, index=False)
+    print(f"CSV file created: {output_csv}")
+
+if __name__ == "__main__":
+    main()
+

discriminator-ep-4.pth

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:83d59c14472ce1cc582798762f7980361e0239e9f524b6b8b6861dab43fd664e
-size 11098603

runs/Jan31_14-52-36_SOM-YT7DYVX-DT/events.out.tfevents.1738363956.SOM-YT7DYVX-DT.46314.0 → discriminator_final.pth

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f7fa2b1c9d0b68193dc5deed1e0f6dc775a2b6a7ae04cd17e0c4268ee4460451
-size 184408
+oid sha256:aa193bfe35924accf8960e23f16ead89275c38839d1d3a0019fa0f3e9176199b
+size 11099058

inputs_example.csv

@@ -0,0 +1,6 @@
+image_uid,age,sex,image_path,split
+0,81,2,/Users/jbrown2/.cache/huggingface/datasets/downloads/extracted/6429865a89f9ae54df1c3c2db5d0f1f25cf7dd43cb87704d76ed08cf8c194aba/OASIS-2/sub-OASIS20133/ses-03/anat/msub-OASIS20133_ses-03_T1w_brain_affine_mni.nii.gz,train
+1,78,2,/Users/jbrown2/.cache/huggingface/datasets/downloads/extracted/6429865a89f9ae54df1c3c2db5d0f1f25cf7dd43cb87704d76ed08cf8c194aba/OASIS-2/sub-OASIS20133/ses-01/anat/msub-OASIS20133_ses-01_T1w_brain_affine_mni.nii.gz,train
+2,87,1,/Users/jbrown2/.cache/huggingface/datasets/downloads/extracted/6429865a89f9ae54df1c3c2db5d0f1f25cf7dd43cb87704d76ed08cf8c194aba/OASIS-2/sub-OASIS20105/ses-02/anat/msub-OASIS20105_ses-02_T1w_brain_affine_mni.nii.gz,train
+3,86,1,/Users/jbrown2/.cache/huggingface/datasets/downloads/extracted/6429865a89f9ae54df1c3c2db5d0f1f25cf7dd43cb87704d76ed08cf8c194aba/OASIS-2/sub-OASIS20105/ses-01/anat/msub-OASIS20105_ses-01_T1w_brain_affine_mni.nii.gz,train
+4,84,1,/Users/jbrown2/.cache/huggingface/datasets/downloads/extracted/6429865a89f9ae54df1c3c2db5d0f1f25cf7dd43cb87704d76ed08cf8c194aba/OASIS-2/sub-OASIS20102/ses-02/anat/msub-OASIS20102_ses-02_T1w_brain_affine_mni.nii.gz,train

requirements.txt

@@ -4,7 +4,7 @@
 torch>=1.12
 
 # MONAI v1.2+ has the 'generative' subpackage with AutoencoderKL, PatchDiscriminator, etc.
-monai>=1.2.0
+monai-weekly
 monai-generative
 
 # For perceptual losses in MONAI's generative module.
@@ -16,4 +16,5 @@ numpy
 nibabel
 tqdm
 tensorboard
-matplotlib
+matplotlib
+datasets