DLUS

GitHub repository for DLUS: Deep Learning-based Segmentation of Prostatic Urethra on Computed Tomography Scans for Treatment Planning

Full paper available at: https://www.sciencedirect.com/science/article/pii/S2405631623000222
Please cite as :
Cubero, L., García-Elcano, L, Mylona, E., Boue-Rafle, A., Cozzarini, C., Ubeira Gabellini, M.G. et al. "Deep learning-based segmentation of prostatic urethra on computed tomography scans for treatment planning." Phys Imaging Radiat Oncol (2023), https://doi.org/10.1016/j.phro.2023.100431.

Any questions, please send an email to Dr. Javier Pascau - [email protected]

https://igt.uc3m.es/

INSTALLATION

Install dependencies:

• Python >= 3.7
• Pytorch
• You should not have any nnU-Net installation in your python environment since DLUS will install its own custom installation.

How to install ?

git clone https://github.com/BSEL-UC3M/DLUS.git
cd DLUS
pip install -e .

0. VARIABLE AND PATH DEFINITIONS

Please, change the following options accordingly:

• ddbb = ... --> Database name. No spaces, no underscore _
• mode = 'dicom' --> Image mode. Options: 'nifti', 'dicom'
• model = 'Mixed_model' --> Model for OAR segmentation. Options: 'FR_model' (French rectum), 'Mixed_model' (French + Italian databases)
• use_manual_OARs = True --> Option to use manual OAR segmentations instead of automatic ones. Options : False, True

1. LOAD ORIGINAL IMAGES

Data must be structured in the following way:
Two directories are needed for each database ddbb:

• input data [data_path] : 'Input' > ddbb
• output data [out_path] : 'Output' > ddbb

Organization of input data: The ddbb folder should contain a different folder for each case to process. In each case folder, the image scan should be saved in a sub-folder named "img", and the manual OAR segmentations - if available - in a sub-folder named "mOAR".

You must create the folder "Input" and add there your database folder. The rest will be done automatically.

As an example:

The loaded images will be saved in 'Output' > ddbb > 'imgs'

• If the manual OARs masks (for the bladder, rectum, prostate, and seminal vesicles) are available and saved in the sub-folder "mOAR", they will also be loaded. To load them in DICOM format, we have included a series of typical names used in the clinic to describe the rectum, bladder, prostate and seminal vesicles. These names can be found in utils.load_data.dicom_to_nifti() and utils.load_data.nifti_data(), and can be updated to include other terminologies by adding them to the available lists. Only the available masks of interest will be saved.

• The loaded mOARs will be saved in 'Output' > ddbb > 'GTs'

2. VOI EXTRACTION

Localization Network + Crop using the centroid of the coarse prosate segmentation.

• IMPORTANT : Check the result to ensure that appropriate VOI has been created !!!!!!!!!!!!!!!. Sometimes some images are not well predicted and it's necessary to modify this VOI manually to ensure that the OARs and urethra segmentations are accurate.

The cropped VOIs will be saved in 'Output' > ddbb > 'VOIs' > 'imagesTs'

• If use_manual_OARs = True --> The manual segmentation of the prostate (if available) will be used directly to create the VOI. For the cases where this mask is not available, the VOI will be created automatically with the previously described method. All the cropped VOIs will be saved in 'Output' > ddbb > 'mVOIs' > 'imagesTs', and the available manual contours in 'Output' > ddbb > 'OARs' > 'manual'

3. OARs SEGMENTATION : Fine Segmentation Network

A trained nnU-Net is called to predict the segmentations of the bladder, rectum, prostate and seminal vesicles. Depending of the model selected (FR_model or Mixed_model), the network called for inference differs. The main distinction between both networks is the protocol to segment the rectum, based on French delineations with FR_model, or trained with joined data from a French and two Italian medical institutions with Mixed_model.

The predicted segmentations will be saved in 'Output' > ddbb > 'OARs' > model

Then, a post-processing step allows to bring the OAR segmentations back to native space and to export them to DICOM format.

The VOI and predicted segmentations in the native space will be saved in 'Output' > ddbb > 'Native'

The VOI and predicted segmentations in DICOM will be saved in 'Output' > ddbb > 'DICOM'

4. DISTANCE MAP COMPUTATION

The bladder and prostate OARs segmentations are used to calculate the distance maps, which are then used to guide the urethra segmentation.

When the automatically generated OARs masks are used, the computed distance maps are saved in 'Output' > ddbb > 'DistanceMaps'

When the manual OARs masks are used (when available), the computed distance maps are saved in 'Output' > ddbb > 'mDistanceMaps'

• We recommend obtaining both the distance maps and running steps 4 and 5 with the manual and automatic contours (use_manual_OARs = True, use_manual_OARs = False) as the automatic ones are smoother in general.

5. URETHRA SEGMENTATION

A trained nnU-Net is called to predict the segmentation of the urethra based on the previously computed distance maps.

The predicted segmentations will be saved in 'Output' > ddbb > 'Urethra' > 'DLUS'

Then, a post-processing step allows to bring the OAR segmentations back to native space and to export them to DICOM format.

The VOI and predicted segmentations in the native space will be saved in 'Output' > ddbb > 'Native'

The VOI and predicted segmentations in DICOM will be saved in 'Output' > ddbb > 'DICOM'

Important: We have noticed that some of the urethra contours extend out of the prostate contour. In those cases, they should be cropped to the borders of the prsotate.