DblurDoseNet

https://github.com/ZongyuLi-umich/DblurDoseNet

PyTorch implementation of DblurDoseNet method for SPECT dosimetry, as described in the paper "DblurDoseNet: A deep residual learning network for voxel radionuclide dosimetry compensating for SPECT imaging resolution", by Zongyu Li, Jeffrey A Fessler, Justin K Mikell, Scott J Wilderman, Yuni K Dewaraja; Medical Physics 49(2):1216-30, Feb. 2021. https://doi.org/10.1002/mp.15397

Dataset

The dataset used for training and testing is available here: https://doi.org/10.7302/ykz6-cn05

Data Configuration

To replicate the paper, you must structure the downloaded dataset correctly. The training script expects a specific folder layout and particular files as input.

1. Directory Structure

After downloading and unzipping the dataset, you should organize the folders as follows. Create a main data directory (e.g., DblurDoseNet_data) and place the train, test, and DVKs folders inside it.

The training data should follow this pattern:

• DblurDoseNet_data/train/petvp01/spect.mat

• DblurDoseNet_data/train/petvp01/density.mat

• DblurDoseNet_data/train/petvp01/doseGT.mat ... and so on for all files in petvp01.

• DblurDoseNet_data/train/petvp04/... ...and so on for all other training phantoms.

The testing data should follow the same pattern:

• DblurDoseNet_data/test/petvp6/spect.mat

• DblurDoseNet_data/test/petvp6/density.mat

• DblurDoseNet_data/test/petvp6/doseGT.mat ...and so on for all files in petvp6.

• DblurDoseNet_data/test/petvp7/... ...and so on for all other testing phantoms.

The DVK kernels should be placed in their own folder:

• DblurDoseNet_data/DVKs/beta.npy
• DblurDoseNet_data/DVKs/gamma.npy

2. Model Input and Output

The DblurDoseNet model is designed to learn a mapping from the low-resolution SPECT image to the high-fidelity ground-truth dose-rate map, using the density map as additional information. For each phantom:

Input: The model takes two .mat files as input:

• spect.mat: The SPECT reconstruction image.
• density.mat: The density map.

Target (Ground Truth): The model learns to produce the doseGT.mat file, which is the ground-truth dose-rate map. The train.py script will load these files from the train directory during the training process.

Training

To train the model, you need to point the training script to your dataset directory. The command below includes a --data_path argument for this purpose (Note: the exact argument name may vary; please check the train.py script if needed).

python3 train.py --data_path [/path/to/DblurDoseNet_data] --batch [batch size] --lr [learning rate] --epochs [# of epochs] <br>

For example, with a batch size of 32, a learning rate of 0.002, and 200 epochs, the training command is:

python3 train.py --data_path ./DblurDoseNet_data --batch 32 --lr 0.002 --epochs 200

Testing

To use the best checkpoint from training (the one with the lowest validation loss), run the following command. Remember to point to your test data as well.

python3 test.py --data_path [/path/to/DblurDoseNet_data] --batch [batch size] --is_best