This repository contains transformer models for the HECKTOR 2022 competition, focusing on the segmentation of head and neck gross tumor volume (GTV). The targets are the primary tumor (GTV-T) and involved lymph nodes (GTV-N).

The following transformer segmentation networks are experimented with on the HECKTOR2022 competition dataset:

• UNETR++
• UNETR
• SWIN Transformer UNet
• nnFormer

The "data_processing" directory contains a collection of scripts for an end-to-end DICOM-to-DICOM pipeline.

• dcm2nii.py: Converts .dcm images and RTSTRUCTS to .nii.gz image and label files. Fuzzy name matching is used to locate both "GTVt" and "GTVn" within the RTSTRUCT file.
• bounding_box.py: Identifies the region of interest within a 192x192x192 bounding box by locating tumors below the brain in the PET image using a thresholding method.
• resample.py: Resamples images (CT, PET) and labels (GTV.nii.gz) to 1x1x1 mm^3 isotropic spacing. GTV-T and GTV-N are stored in a single .nii.gz file, with values of 1 for T and 2 for N.
• resample_ct_test.py: Similar to resample.py but solely resamples CT images. Also include the funciont to revert resample based on the bounding-box information.
• copy_data_for_nnUNet_Task005_CT.py: A script for data conversion and copying that follows the nnUNet format.
• nii2dcm.py: Converts .nii.gz GTV files back to .dcm RTSTRUCT files with the naming convention "AI_GTV_T" and "AI_GTV_N". Both names are rendered in yellow.