Automatic segmentation of dentate nuclei

This method is implemented in Matlab19a using the Deep Learning Toolbox. The CNN implemented in 2D is an automated segmentation method using the non-DWI (b0) images from a DWI dataset. CNN training was performed on data from healthy subjects with high resolution from the Human Connectome Project (HCP). To demonstrate its clinical applicability, the developed CNN was further used to segment the DNs of a subset of subjects affected by Temporal Lobe Epilepsy (TLE).

For more details, see the original publication: Automatic Segmentation of Dentate Nuclei for Microstructure Assessment: Example of Application to Temporal Lobe Epilepsy Patients, DOI: 10.1007/978-3-030-73018-5_21.

If you use this code in your work, please cite the above paper.

N.B. put path with single comma ' ' (example: '/media/bcc/bcc-data/MARTA/HCP')

Three functions: resampling_normalize.m, CNN_postprocessing.m and segment_DN_SUIT.m

---

1) resampling_normalize(dir_seg, ref_img)

Resampling and normalize the b0 image:

resampling image (that you want to segment) to spatial resolution of HCP (this resolution is that was used for CNN training)

intensity normalization: mean=0 and std=1 for the voxels that belong to brain -> it is required a mask of the brain

REQUIRED:

• flirt (FSL)
• image HCP (as refrence for resampling) "SIGNAL.nii.gz" (available in Directory "download")

INPUT:

• dir_seg: directory with content N directory as N subjects to segment. In each directory there are: b0 ("b0.nii.gz") and mask of brain ("brain_mask.nii.gz").
• ref_img: path to b0 of HCP to use as reference; path must end with "/SIGNAL.nii.gz"

OUTPUT:

• b0 resampled, name "b0_125.nii.gz"
• b0 resampled and normalized, name "B0_N.nii"

---

2) CNN_postprocessing(dir_seg, path_CNN)

Segment DN using CNN and clear FP using segmentation DN obteined with SUIT

REQUIRED:

• DN segmented with SUIT: "dentati_suit.nii.gz" (if you have not use script "segment_DN_SUIT.m")
• fsl (fslmath for dilatate DN of SUIT)
• check to have matRead.m and dicePixelClassificationLayer.m (available in Directory "download")

INPUT:

• dir_seg: path to directory that contain images that you want to segment
• path_CNN: path to Directoy download where are "rete.mat"

OUTPUT:

• "DN_CNN" segmentation obtained with CNN to your resolution

---

3) segment_DN_SUIT(dir_seg)

DN segmentation with ATLAS SUIT

REQUIRED:

• fsl: for isolate DN from atlas
• SUIT: in SPM toolbox
• "T1.nii": T1 image (this must be registred to b0 image)
• "suit_isolate_seg_gio.m" same script of suit_isolate_seg but return bb_box coordintate of crop (available in Directory "download")

INPUT:

• dir_seg: path to directory that contain images that you want to segment, In this directory there are N directory as N subjects to segment. In each of N directory there are b0 and T1
• suit_atlas: path to Cerebellum-SUIT.nii.gz (available in Directory "download") this path have to end with "/Cerebellum-SUIT.nii.gz"

OUTPUT:

• "seg_den_suit.nii": DN segmentation obtain with atlas SUIT

---

CNN Architecture