Prostate Segmentation according to the PI-RADS standard using a 3D-CNN

Welcome, to the GitHub page of my Master Thesis Project. This page will provide links to the Thesis report, the software HERO - for which the model will be implemented in the free version, as well as a link to download the trained model as a frozen graph (TensorFlow), ready to use in Python.

The Thesis aimed to create an automatic solution for segmentation of the prostate and its internal anatomical zones from T2 MR images, as this can be a very time-consuming task if performed manually. The result is this 3D-CNN model.

---

HERO

In the autumn of 2022, the model will be available in the free version of HERO (www.heroimaging.com) - a graphical programming platform that enables complex image analysis, in an easy and interactive way.

This will be the easiest and most straight-forward way of using the model.

The model takes a T2 MRI prostate (DICOM) image as input and returns a segmented multilabel (DICOM) image with the anatomical zones as output - which can also easily be separated into individual RTSTRUCTs.

To try the model on a larger dataset (n=98), one can use the PROSTATEx dataset from the Cancer Imaging Archive, avaliable at:

https://wiki.cancerimagingarchive.net/pages/viewpage.action?pageId=70230177

This is the training dataset, but the segmentations have been modified in the trainig of the model according to the description in the Thesis.

Frozen graph - Python

The trained model is also available as a frozen graph (from TensorFlow), which can be used directly in Python.

This setup lacks the preprocessing and postprocessing included in the HERO workflow, and thus the output will be six individual images representing each individual zone (PZ, CZ, TZ, Urethra, AFS, and Background). If you are interested in the Python code for the pre- and postprocessing to try it out before the model is avaliable in HERO, you can reach out at: william.holmlund@umu.se.

The frozen graph model is available at:

https://nonpi.s3.eu-north-1.amazonaws.com/mtk/resources/model-zoo/0005.pb

---

For the interested reader that would like to know more about the project, the full Thesis report is avaliable to download as a pdf at:

https://www.diva-portal.org/smash/record.jsf?dswid=-4000&pid=diva2%3A1654535&c=2&searchType=SIMPLE&language=sv&query=william+holmlund&af=%5B%5D&aq=%5B%5B%5D%5D&aq2=%5B%5B%5D%5D&aqe=%5B%5D&noOfRows=50&sortOrder=author_sort_asc&sortOrder2=title_sort_asc&onlyFullText=false&sf=all