Awesome

Head CT preprocess

This is a pipeline for preprocessing computed tomography (CT) head scans. It converts them from DICOM format to NIfTI, corrects their orientation and position, removes non brain tissue and performs HU windowing.

This pipeline works for NCCT (non contrast computed tomography) and CTA (computed tomography angiography) scans both.

test.py

Don't now where to start? Try downloading and unziping the Patient_Reports_16.zip from The UCLH Stroke EIT Dataset. Then, try to run the simple test.py script (don't forget to update de to_scan variable first). This will perform the preprocessing you see on the example images above and below.

These examples images are from the Series 005 [CT - Thin Bone 1 0 Bone Sharp] DICOM directory from the UCLH Stroke EIT Dataset (Goren, Nir; Dowrick, Thomas, Avery, James; Holder, David).

convert_and_preprocess.py

This program is responsible for the following:

1. Conversion from DICOM to NIfTI, using dcm2niix.
2. Registration to a common brain atlas. This step normalizes the scans so they have the same position and orientation of a brain scan given as refference, using the FLIRT tool from FSL. The refferences used are the MNI152_T1 with a thickness that depends on the thickness of the input scan.
3. Remove non brain tissue. This step is done using an adapted version of the algoritm proposed in Validated automatic brain extraction of head ct images, NeuroImage (2015) by Muschelli, Ullman, Mould, et al. that uses the BET tool from FSL. The algorithm proposed in this article was designed for NCCTs only. To apply it to CTA images, the first windowing step was updated to use a window in the range of [-75, 425].

You need to change the variable to_scan to specify the source DICOMs directory. This directory should have a layout similar to this:

to_scan
├── idProcessoLocal-1
│   ├── DICOM
│   │   └── 0000662F
│   │       └── AA923232
│   │           └── AA331FC9
│   │               ├── 00001AFC    # DICOM dir 1 of patient idProcessoLocal-1
│   │               │   ├── EE00D32D
│   │               │   ├── EE01CD46
│   │               │  ...
│   │               ├── 000031E3    # DICOM dir 2 of patient idProcessoLocal-1
│   │               │   ├── EE011BBA
│   │               │   ├── EE0CFE6F
......             ... ...
│   ├── README.TXT
│   └── SECTRA
│       └── CONTENT.XML
...
└── idProcessoLocal-N
    ├── DICOM
    │   └── 00002B78
    │       └── AAA1DB91
    │           └── AAFA3E4F
    │               ├── 00003FA1    # DICOM dir 1 of patient idProcessoLocal-N
    │               │   ├── EE02CAD0
    │               │   ├── EE036E90
   ...             ... ...
    ├── README.TXT
    └── SECTRA
        └── CONTENT.XML

Note: the convert_and_preprocess.py requires this very specific directory structure because this is the structure of the DICOM files I exported from a SECTRA software program. If you have a different file structure, take a look at the script test.py.

cta_bet.py

Further CTA skull strip. Some CTAs had some non brain tissue that persisted after step 3. These exams can be further cleaned by using a brain mask provided by the NCCT exam of the same patient.

window_HU.py

HU range windowing. NCCTs are normalized to be in the range [0,100] and CTAs to be in the range [0,200].

Disclaimer

I am not an expert in medical imaging. From what I understand, DICOM headers vary from brand to brand, so the is_CT and is_CTA functions may not work as intended for your scans.

For the CT images I had, 2mm thickness for the NCCT scans and 1mm thickness for the CTA scans worked the best. I am not sure if this pipeline produces decent results for other thickness values.

The pipeline is not perfect: sometimes the scans end up in the wrong orientation or with holes.

License

These programs are licensed under the MIT License - see the LICENSE file for details.