The code below uses a trained Denoising Diffusion Probabilistic Models (DDPM) model to generate a large batch of 2D MR image samples that are saved as a numpy array. These are also referred to as Stochastic Object Models (SOMs) in the rest of the implementation.

conda activate ddpm
cd src/demo1
# Python file and its input and outputs -----------------------------------------------------------------------------
PY_FILE=scripts/image_sample_newdataset2_centercrop.py
MODEL_PATH=trained_DDPM_model/ema_0.9999_1100000.pt
OUTPUT_FLD="test_out_200k"

# Important ddpm parameters --------------------------
dSTEP=1000
nSAMPLES=1
bSIZE=1

# CMD_ARGUMENTS aligning with how we trained the DDPM model -----------------------------------------------------------------------------
MODEL_FLAGS="--image_size 384 --attention_resolutions 32,16,8 --num_channels 128 --num_head_channels 64 --num_res_blocks 2 --resblock_updown True --use_scale_shift_norm True --learn_sigma True"
DIFFUSION_FLAGS="--diffusion_steps ${dSTEP} --noise_schedule cosine "
sample_FLAGS="--save_dir ${OUTPUT_FLD}/HCP_brain_384x384_cropped_260x311_step1100k_ema_samples/ --num_samples ${nSAMPLES} --batch_size ${bSIZE}"

time python ${PY_FILE} --model_path ${MODEL_PATH} $MODEL_FLAGS ${DIFFUSION_FLAGS} ${sample_FLAGS}

The generated .npz file is the object input for demos 2 and 3. With the default example above, the handoff file is:

test_out_200k/HCP_brain_384x384_cropped_260x311_step1100k_ema_samples/samples_10000x260x311x1.npz

1. The DDPM¹ was trained using the HCP's young adult dataset² :

   • This dataset consists of 1,113 subjects scanned on a customized Siemens 3T MRI system.
   • From each patient, 10 axial slices within their Cerebrospinal fluid (CSF) regions were extracted to train the DDPM model.

2. The no. of diffusion steps is an important parameter when running this ddpm-based MRI data generation code.

   
   MR images for different diffusion steps

3. Appropriately validate the generated data using metrics relevant to your clinical application. Below are the image statistics corresponding to the mean, standard deviation, white matter area-to-intracranial area ratio, and gray matter area-to-intracranial area ratio, computed using 1,000 real and 1,000 DDPM-generated MR images at each step.

   
   Image statistics for different diffusion steps

4. Approximately 217,000 images were generated, including 168,000 for DLMO training, 40,000 for UNet training, 8,000 for image fidelity and AUC analysis, and 640 × 2 for the 2AFC validation study of the DLMO.