prostate_t2map

Code to develop and evaluate NNs for parameter estimation

This archive contains all the code and models used in the following manuscript: Improved Quantitative Parameter Estimation for Prostate T2 Relaxometry using Convolutional Neural Networks https://doi.org/10.1101/2023.01.11.23284194

You can use this code to generate the synthesized data, train the models, run inference, and create all the figures in the paper. However, this process is not fully automated - you will need to do this in parts, with some manual steps. You can use the pre-trained models, or retrain them youself with a few days of compute time.

1. Configure python. You can use the requirements.txt file to create a python virtual environment matching mine. I have never been able to get that to fully install an environment correctly, but it is a good starting point. I ran all of the code for this paper from the Spyder development environment, and for that I usually have to install a specific older verion (5.2.2) then upgrade to the latest, because there are missing packagesin the later versions.

2. Download ImageNet data. You need to download the full ImageNet dataset (165GB), even though we only used the validation dataset (50k images, 6.4GB) for this paper. Go to this website: https://www.kaggle.com/competitions/imagenet-object-localization-challenge/data agree to the ImageNet competition rules and download the full dataset. The validation data are in ILSVRC/Data/CLS-LOC/val/, with first file ILSVRC2012_val_00000001.JPEG (a sea snake on a beach)

3. Update paths. In the python file utility_functions.py I have several hardwired paths, either relative to my "base" path (/home/pbolan/prostate_t2map), or pointing to a location with the downloaded imagenet data. You'll want to update these for your system.

4. Download and organize the in vivo data. Under the base directory create a "datasets" directory. The publicly available files are available here: https://conservancy.umn.edu/handle/11299/243192 as a single file, images.tgz. Expand this, and you'll find imagesTr (training) and imagesTs (testing). You'll first move these into folders to create three 3D datasets (set1 = development, set2=training, set3=testing).

Create a folder called $BASE/datasets/invivo_set1/images/, and copy the first nine training files (prostate_00*.nii.gz) from imagesTr into there. This is optional. Create a folder called $BASE/datasets/invivo_set2/images/, and move all the *.nii.gz files from imagesTr into there. Create a folder called $BASE/datasets/invivo_set3/images/, and move all the *.nii.gz files from imagesTs into there.

All of those files are 3D niftis, and they need to be extracted into 2D slices. To do this, run the script extract_slices_from_invivo_datasets.py.

5. Generate synthetic data. Run build_synthetic_datasets.py to build the testing and training datasets. Note I have a big (10k) training set, and a smaller one (1k) that I used for development.

6. Train models. Run train_1d.py and train_cnn.py. These will take a few days, depending on your system, and will write out all the models into $BASE/models/*.pt. You can skip this and just use the pre-trained models available on github.

7. Perform inference. In the file inference.py, the inference is broken into 3 parts: part A: synthetic data; part B: inviv; part C: invivo with noise addition; You can run inference.py to do all of them, or just run the parts you want. This will create the predicted values in the folder $BASE/predictions, about 7GB.

8. Run analyses and generate plots. This is more manual, and done in several steps. While developing I often used inline graphics (%matplotlib inline), but for the paper I used Qt as the renderer, saving png and svg files and arranging them into figures manually. make_demo_figure.py will make the plots for figure 1, and the top section for figures 3 and S1. You'll need to run several times, changing the switch value to make all plots. plot_example_partA.py will make the plots for figure 3. make_plots_partA.py will do all per-slice and per-pixel evaluations on synthetic data, figures 4 and 5. Will take a few hours for the per-pixel stuff. plot_example_partB.py will make plots for figure 6. plot_example_partC.py will make figure 7. evaluate_partC_byslice.py will make figure 8.