Add phase image as input channel to improve segmentation performance

[jcohenadad]

Description
Currently, the model uses only the magnitude image for input. Incorporating the phase image as an additional input channel may provide complementary information, particularly near tissue boundaries or susceptibility-affected regions, potentially improving segmentation accuracy.

Proposed Change

• Modify the data loading pipeline to support two-channel input: magnitude + phase.
• Retrain model and compare with magnitude-only baseline.

Rationale
Preliminary observations suggest that phase contrast enhances anatomical delineation in our target regions (e.g., GM/WM interface, lesions), which could lead to improved pseudo-Dice scores.

Impact

• Minimal code change in preprocessing and model input layers.
• May require re-tuning of normalization and training schedule.

[jcohenadad]

Using prototype code in 721ef04:

python extract_slices.py --path-data ${PATH_DATA} --label-folder derivatives/labels_combined-sc-gm --labels combined --path-out ${PATH_PROCESSED}/data_slice --adjacent-slices 1 --use-phase
python convert_bids_to_nnunet.py --path-data ${PATH_PROCESSED}/data_slice --label-json ${PATH_DATA}/derivatives/labels_combined-sc-gm/sub-01/anat/sub-01_part-mag_chunk-02_T2starw_label-combined_classes.json --tasknumber 502 --adjacent-slices 1 --use-phase

Here is the dataset.json:

{
    "channel_names": {
        "0": "T2starw_offset0_mag",
        "1": "T2starw_offset0_phase",
        "2": "T2starw_offset-1_mag",
        "3": "T2starw_offset-1_phase",
        "4": "T2starw_offset1_mag",
        "5": "T2starw_offset1_phase"
    },
    "labels": {
        "background": 0,
        "SC_without_GM": 1,
        "SC_GM": 2
    },
    "numTraining": 9,
    "file_ending": ".nii.gz",
    "licence": "CC-BY 4.0",
    "converted_by": "Please enter your name, especially when sharing datasets with others in a common infrastructure!",
    "name": "Segmentation",
    "reference": "Julien Cohen-Adad, Postmortem SC MRI",
    "release": "1.0.0",
    "overwrite_image_reader_writer": "NibabelIOWithReorient"
}

[jcohenadad]

Below comparison with/without the phase, using 3 adjacent slices:

Using the phase yields promising results (esp. because the number of epochs for the phase model was 230 vs. 520 without the phase). Further investigations are required.