Detection of local prostate cancer recurrence from PET/CT scans using deep learning

This repository contains the accompanying code for the article:

Korb M, Efetürk H, Jedamzik T, Hartrampf PE, Kosmala A, Serfling SE, Michalski K, Dirk R, Buck AK, Werner RA, Schlötelburg W, and Ankenbrand M. Detection of local prostate cancer recurrence from PET/CT scans using deep learning. In preparation

Important

Large outputs (e.g. model weights) are deposited on Zenodo (10.5281/zenodo.15174580). Training data can not be shared publically. To understand the structure of the data, these files are included as symbolic links that point outside of the repository.

Pre-processing

Nifti conversion

Exported and pseudonomized PET and CT dicom images for each examination were converted to nifti format using dcm2niix (Chris Rorden's dcm2niiX version v1.0.20220720 (JP2:OpenJPEG) (JP-LS:CharLS) GCC5.5.0 x86-64 (64-bit Linux)). Those niftis are saved in data/nifti (train and validation set) and data/nifti_ts2024 (test set).

Prostate segmentation

The prostate and urinary bladder were segmented with TotalSegmentator (version 2.1.0) in all ct images.

for i in data/nifti*/*_ct.nii.gz
do
	TotalSegmentator -i $i -o analysis/totalsegmentator2/$(basename $i _ct.nii.gz) -rs prostate urinary_bladder
done

Cropping around prostate (or urinary bladder)

Cropping a 20x20x20 cm³ cube around the centroid of the prostate (if detected) or urinary bladder (otherwise) with the script code/preprocessing/crop_by_prostate_or_ub.py. The cropped files are saved in data/cropped_nifti.

Conversion from BQML to SUV

Determine factors for SUV conversion

The converted nifti PET files have values in the BQML unit. In order to convert them to SUV, individual conversion factors have to be determined. This was done by applying code/preprocessing/bqml_to_suv.py to the dicom files (containing the relevant header information) to create the factors in data/suv_factors.tsv.

Convert cropped PET niftis

while read pid suv
do
	fslmaths data/cropped_nifti/${pid}_pet.nii.gz -mul $suv data/cropped_nifti_suv/${pid}_pet.nii.gz
done <<(tail -n +2 data/suv_factors.tsv)

Get ranges for scaling

python code/preprocessing/nii_range.py data/cropped_nifti/*_ct.nii.gz >analysis/cropped_ct_range.tsv
python code/preprocessing/nii_range.py data/cropped_nifti_suv/*_pet.nii.gz >analysis/cropped_pet_suv_range.tsv