OSCR

Learning-based segmentation of diffusion-weighted MR images with arbitrary <i>q</i>-space samplings.

Code ↔ Paper

11 matches between paragraphs of the paper and lines of its authors' code, computed by the harvester (lexical-v1). Click a colored paragraph or line to see its counterpart.

The 11 matches · 2 of them tie a paragraph to a whole file, not to given lines: weak matches, whose lines are not tinted
  1. [1] § Materials and Methods › Data › Preprocessing › Reference segmentations ↔ DISCUSseg/data_loader/dkt_labels.py, lines 1–23 · score 0.69 · Desikan Killiany Tourville, DKT, FreeSurfer, mapping
  2. [2] § Materials and Methods › Evaluation › Labels ↔ data_loader/load_neuroimaging_data.py, lines 405–520 · score 0.69 · corpus callosum, cortical region, hemisphere, hypointensities, lateral, ventricles
  3. [3] § Materials and Methods › Method › Architecture › Details ↔ DISCUSseg/run_prediction.py, lines 178–297 · score 0.63 · brain mask, DISCUS encoder, sagittal, axial, memory, inference
  4. [4] § Materials and Methods › Method › Architecture › Details ↔ DISCUSseg/data_loader/discus_preprocessing.py, the whole file · a weak match · score 0.63 · brain mask, DISCUS encoder, inside, width, inference, volume
  5. [5] § Materials and Methods › Method › Training › Phase 1: DISCUS pre-training ↔ DISCUSseg/train.py, lines 171–242 · score 0.62 · DISCUS reconstructs, train DISCUS, DISCUS encoder, decoder, loss, signals
  6. [6] § Materials and Methods › Method › Training › Phase 1: DISCUS pre-training ↔ DISCUSseg/data_loader/discus_preprocessing.py, the whole file · a weak match · score 0.59 · DISCUS reconstructs, DISCUS encoder, subsets, decoder, loss, signals
  7. [7] § Materials and Methods › Reference methods › SynthSeg ↔ process.sh, lines 1–59 · score 0.54 · fractional anisotropy, diffusion tensor, FA, fitted, brain, mask
  8. [8] § Materials and Methods › Data › Preprocessing › Masks ↔ DISCUSseg/run_prediction.py, lines 319–407 · score 0.54 · brain mask, FSL, binary, probability, inference, maps
  9. [9] § Materials and Methods › Reference methods › DDParcel ↔ process.sh, lines 1–59 · score 0.52 · fractional anisotropy, diffusion tensor, eigenvalues, brain, masking, .2
  10. [10] § Materials and Methods › Method › Training › Alternative training schemes › Multi-task training for segmentation and reconstruction ↔ DISCUSseg/train.py, lines 171–242 · score 0.52 · DISCUS encoder, rotation, embeddings, reconstruction, decoder, augmented
  11. [11] § Materials and Methods › Data › Preprocessing › Reference segmentations ↔ data_loader/load_neuroimaging_data.py, lines 405–520 · score 0.51 · corpus callosum, vessels, DKT, cortical, mapping, segmentations

Paper

Loaded from Europe PMC by your browser, not stored by OSCR: doi.org · Europe PMC

The paper is loaded when this pane is shown.

The authors' code

Python · 987 lines · 36 KB · Unlicense · 2 matches

  2. # Copyright 2019 Image Analysis Lab, German Center for Neurodegenerative Diseases (DZNE), Bonn
  3. #
  4. # Licensed under the Apache License, Version 2.0 (the "License");
  5. # you may not use this file except in compliance with the License.
  6. # You may obtain a copy of the License at
  7. #
  8. # http://www.apache.org/licenses/LICENSE-2.0
  9. #
  10. # Unless required by applicable law or agreed to in writing, software
  11. # distributed under the License is distributed on an "AS IS" BASIS,
  12. # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13. # See the License for the specific language governing permissions and
  14. # limitations under the License.
  17. # IMPORTS
  18. import nibabel as nib
  19. import numpy as np
  20. import h5py
  21. import scipy.ndimage.morphology as morphology
  22. import scipy.ndimage as ndimage
  23. import scipy.ndimage.filters as filters
  24. import sys
  25. import glob, os
  27. from skimage.measure import label
  28. from torch.utils.data.dataset import Dataset
  29. from .conform import is_conform, conform, check_affine_in_nifti
  31. ##
  32. # Helper Functions
  33. ##
  36. # Conform an MRI brain image to UCHAR, RAS orientation, and 1mm isotropic voxels
  37. def load_and_conform_image(img_filename, interpol=1, logger=None, imagetype='image'):
  38. """
  39. Function to load MRI image and conform it to UCHAR, RAS orientation and 1mm isotropic voxels size
  40. (if it does not already have this format)
  41. :param str img_filename: path and name of volume to read
  42. :param int interpol: interpolation order for image conformation (0=nearest,1=linear(default),2=quadratic,3=cubic)
  43. :return:
  44. """
  45. orig = nib.load(img_filename)
  47. if not is_conform(orig):
  49. if logger is not None:
  50. logger.info('Conforming image to UCHAR, RAS orientation, and 1mm isotropic voxels')
  51. else:
  52. print('Conforming image to RAS orientation, and 1mm isotropic voxels')
  54. if len(orig.shape) > 3 and orig.shape[3] != 1:
  55. sys.exit('ERROR: Multiple input frames (' + format(orig.shape[3]) + ') not supported!')
  57. # Check affine if image is nifti image
  58. if img_filename[-7:] == ".nii.gz" or img_filename[-4:] == ".nii":
  59. if not check_affine_in_nifti(orig, logger=logger):
  60. sys.exit("ERROR: inconsistency in nifti-header. Exiting now.\n")
  62. # conform
  63. orig = conform(orig, interpol, imagetype=imagetype)
  65. # Collect header and affine information
  66. header_info = orig.header
  67. affine_info = orig.affine
  68. orig = np.asarray(orig.get_fdata(), dtype=np.float32)
  70. return header_info, affine_info, orig
  72. def cropparameters(orig_, img_filename):
  74. data = orig_.get_fdata()
  78. # Transformation for mapping
  79. def transform_axial(vol, coronal2axial=True):
  80. """
  81. Function to transform volume into Axial axis and back
  82. :param np.ndarray vol: image volume to transform
  83. :param bool coronal2axial: transform from coronal to axial = True (default),
  84. transform from axial to coronal = False
  85. :return:
  86. """
  87. if coronal2axial:
  88. return np.moveaxis(vol, [0, 1, 2], [1, 2, 0])
  89. else:
  90. return np.moveaxis(vol, [0, 1, 2], [2, 0, 1])
  93. def transform_sagittal(vol, coronal2sagittal=True):
  94. """
  95. Function to transform volume into Sagittal axis and back
  96. :param np.ndarray vol: image volume to transform
  97. :param bool coronal2sagittal: transform from coronal to sagittal = True (default),
  98. transform from sagittal to coronal = False
  99. :return:
  100. """
  101. if coronal2sagittal:
  102. return np.moveaxis(vol, [0, 1, 2], [2, 1, 0])
  103. else:
  104. return np.moveaxis(vol, [0, 1, 2], [2, 1, 0])
  107. # Thick slice generator (for eval) and blank slices filter (for training)
  108. def get_thick_slices(img_data, slice_thickness=3):
  109. """
  110. Function to extract thick slices from the image
  111. (feed slice_thickness preceeding and suceeding slices to network,
  112. label only middle one)
  113. :param np.ndarray img_data: 3D MRI image read in with nibabel
  114. :param int slice_thickness: number of slices to stack on top and below slice of interest (default=3)
  115. :return:
  116. """
  117. h, w, d = img_data.shape
  118. img_data_pad = np.expand_dims(np.pad(img_data, ((0, 0), (0, 0), (slice_thickness, slice_thickness)), mode='edge'),
  119. axis=3)
  120. img_data_thick = np.ndarray((h, w, d, 0), dtype=np.uint8)
  122. for slice_idx in range(2 * slice_thickness + 1):
  123. img_data_thick = np.append(img_data_thick, img_data_pad[:, :, slice_idx:d + slice_idx, :], axis=3)
  125. return img_data_thick
  128. def filter_blank_slices_thick(img_vol, label_vol, weight_vol, threshold=50):
  129. """
  130. Function to filter blank slices from the volume using the label volume
  131. :param np.ndarray img_vol: orig image volume
  132. :param np.ndarray label_vol: label images (ground truth)
  133. :param np.ndarray weight_vol: weight corresponding to labels
  134. :param int threshold: threshold for number of pixels needed to keep slice (below = dropped)
  135. :return:
  136. """
  137. # Get indices of all slices with more than threshold labels/pixels
  138. select_slices = (np.sum(label_vol, axis=(0, 1)) > threshold)
  140. # Retain only slices with more than threshold labels/pixels
  141. img_vol = img_vol[:, :, select_slices, :]
  142. label_vol = label_vol[:, :, select_slices]
  143. weight_vol = weight_vol[:, :, select_slices]
  145. return img_vol, label_vol, weight_vol
  148. # weight map generator
  149. def create_weight_mask(mapped_aseg, max_weight=5, max_edge_weight=5):
  150. """
  151. Function to create weighted mask - with median frequency balancing and edge-weighting
  152. :param mapped_aseg:
  153. :param max_weight:
  154. :param max_edge_weight:
  155. :return:
  156. """
  157. unique, counts = np.unique(mapped_aseg, return_counts=True)
  159. # Median Frequency Balancing
  160. class_wise_weights = np.median(counts) / counts
  161. class_wise_weights[class_wise_weights > max_weight] = max_weight
  162. (h, w, d) = mapped_aseg.shape
  164. weights_mask = np.reshape(class_wise_weights[mapped_aseg.ravel()], (h, w, d))
  166. # Gradient Weighting
  167. (gx, gy, gz) = np.gradient(mapped_aseg)
  168. grad_weight = max_edge_weight * np.asarray(np.power(np.power(gx, 2) + np.power(gy, 2) + np.power(gz, 2), 0.5) > 0,
  169. dtype=np.float32)
  171. weights_mask += grad_weight
  173. return weights_mask
  176. # class unknown filler (cortex)
  177. def fill_unknown_labels_per_hemi(gt, unknown_label, cortex_stop):
  178. """
  179. Function to replace label 1000 (lh unknown) and 2000 (rh unknown) with closest class for each voxel.
  180. :param np.ndarray gt: ground truth segmentation with class unknown
  181. :param int unknown_label: class label for unknown (lh: 1000, rh: 2000)
  182. :param int cortex_stop: class label at which cortical labels of this hemi stop (lh: 2000, rh: 3000)
  183. :return:
  184. """
  185. # Define shape of image and dilation element
  186. h, w, d = gt.shape
  187. struct1 = ndimage.generate_binary_structure(3, 2)
  189. # Get indices of unknown labels, dilate them to get closest sorrounding parcels
  190. unknown = gt == unknown_label
  191. unknown = (morphology.binary_dilation(unknown, struct1) ^ unknown)
  192. list_parcels = np.unique(gt[unknown])
  194. # Mask all subcortical structures (fill unknown with closest cortical parcels only)
  195. mask = (list_parcels > unknown_label) & (list_parcels < cortex_stop)
  196. list_parcels = list_parcels[mask]
  198. # For each closest parcel, blur label with gaussian filter (spread), append resulting blurred images
  199. blur_vals = np.ndarray((h, w, d, 0), dtype=np.float32)
  201. for idx in range(len(list_parcels)):
  202. aseg_blur = filters.gaussian_filter(1000 * np.asarray(gt == list_parcels[idx], dtype=np.float32), sigma=5)
  203. blur_vals = np.append(blur_vals, np.expand_dims(aseg_blur, axis=3), axis=3)
  205. # Get for each position parcel with maximum value after blurring (= closest parcel)
  206. unknown = np.argmax(blur_vals, axis=3)
  207. unknown = np.reshape(list_parcels[unknown.ravel()], (h, w, d))
  209. # Assign the determined closest parcel to the unknown class (case-by-case basis)
  210. mask = gt == unknown_label
  211. gt[mask] = unknown[mask]
  213. return gt
  215. # class unknown filler (cortex)
  216. def fill_WMhyper_per_hemi(gt, WMhyper_label=77, replace_labels=[2, 41]):
  217. """
  218. Function to replace label 1000 (lh unknown) and 2000 (rh unknown) with closest class for each voxel.
  219. :param np.ndarray gt: ground truth segmentation with class unknown
  220. :param int WMhyper_label: class label for unknown: 77
  221. :param int cortex_stop: class label at which cortical labels of this hemi stop (lh: 2000, rh: 3000)
  222. :return:
  223. """
  224. # Define shape of image and dilation element
  225. h, w, d = gt.shape
  226. struct1 = ndimage.generate_binary_structure(3, 2)
  228. # Get indices of unknown labels, dilate them to get closest sorrounding parcels
  229. unknown = gt == WMhyper_label
  230. if np.sum(unknown) == 0:
  231. return gt
  233. unknown = (morphology.binary_dilation(unknown, struct1) ^ unknown)
  234. list_parcels = np.unique(gt[unknown])
  236. if np.intersect1d(list_parcels, replace_labels).shape[0] > 0:
  237. list_parcels = np.intersect1d(list_parcels, replace_labels)
  239. # For each closest parcel, blur label with gaussian filter (spread), append resulting blurred images
  240. blur_vals = np.ndarray((h, w, d, 0), dtype=np.float32)
  242. for idx in range(len(list_parcels)):
  243. aseg_blur = filters.gaussian_filter(1000 * np.asarray(gt == list_parcels[idx], dtype=np.float32), sigma=5)
  244. blur_vals = np.append(blur_vals, np.expand_dims(aseg_blur, axis=3), axis=3)
  246. # Get for each position parcel with maximum value after blurring (= closest parcel)
  247. unknown = np.argmax(blur_vals, axis=3)
  248. unknown = np.reshape(list_parcels[unknown.ravel()], (h, w, d))
  250. # Assign the determined closest parcel to the unknown class (case-by-case basis)
  251. mask = gt == WMhyper_label
  252. gt[mask] = unknown[mask]
  254. return gt
  256. # Label mapping functions (to aparc (eval) and to label (train))
  257. def map_label2aparc_aseg(mapped_aseg):
  258. """
  259. Function to perform look-up table mapping from label space to aparc.DKTatlas+aseg space
  260. :param np.ndarray mapped_aseg: label space segmentation (aparc.DKTatlas + aseg)
  261. :return:
  262. """
  263. aseg = np.zeros_like(mapped_aseg)
  264. labels = np.array([0, 2, 4, 5, 7, 8, 10, 11, 12, 13, 14,
  265. 15, 16, 17, 18, 24, 26, 28, 31, 41, 43, 44,
  266. 46, 47, 49, 50, 51, 52, 53, 54, 58, 60, 63,
  267. 192, 1001, 1002, 1003, 1005, 1006, 1007, 1008, 1009, 1010, 1011,
  268. 1012, 1013, 1014, 1015, 1016, 1017, 1018, 1019, 1020, 1021, 1022,
  269. 1023, 1024, 1025, 1026, 1027, 1028, 1029, 1030, 1031, 1032, 1033, 1034, 1035,
  270. 2002, 2005, 2010, 2012, 2013, 2014, 2016, 2017, 2021, 2022, 2023,
  271. 2024, 2025, 2028])
  273. h, w, d = aseg.shape
  275. aseg = labels[mapped_aseg.ravel()]
  277. aseg = aseg.reshape((h, w, d))
  279. return aseg
  282. def map_aparc_aseg2label(aseg, aseg_nocc=None):
  283. """
  284. Function to perform look-up table mapping of aparc.DKTatlas+aseg.mgz data to label space
  285. :param np.ndarray aseg: ground truth aparc+aseg
  286. :param None/np.ndarray aseg_nocc: ground truth aseg without corpus callosum segmentation
  287. :return:
  288. """
  289. aseg = aseg.astype(np.int16)
  291. aseg_temp = aseg.copy()
  292. aseg[aseg == 80] = 77 # Hypointensities Class
  293. aseg[aseg == 85] = 0 # Optic Chiasma to BKG
  294. aseg[aseg == 62] = 41 # Right Vessel to Right WM
  295. aseg[aseg == 30] = 2 # Left Vessel to Left MW
  296. aseg[aseg == 72] = 24 # 5th Ventricle to CSF
  298. # Fan: change wm-lh wm-rh to cerebral WM
  299. aseg[(aseg_temp >= 3000) & (aseg_temp < 3999)] = 2
  300. aseg[(aseg_temp >= 4000) & (aseg_temp < 4999)] = 41
  302. aseg[aseg_temp == 5001] = 2
  303. aseg[aseg_temp == 5002] = 41
  305. # Fan: add CC
  306. aseg[(aseg >= 251) & (aseg <= 255)] = 251
  308. # If corpus callosum is not removed yet, do it now
  309. if aseg_nocc is not None:
  310. cc_mask = (aseg >= 251) & (aseg <= 255)
  311. aseg[cc_mask] = aseg_nocc[cc_mask]
  313. aseg[aseg == 3] = 0 # Map Remaining Cortical labels to background
  314. aseg[aseg == 42] = 0
  316. # If ctx-unknowns are not filled yet, do it now
  317. if np.any(np.in1d([1000, 2000], aseg.ravel())):
  318. aseg = fill_unknown_labels_per_hemi(aseg, 1000, 2000)
  319. aseg = fill_unknown_labels_per_hemi(aseg, 2000, 3000)
  321. # Fan: Make all right to be left:
  322. cortical_label_mask = (aseg >= 2000) & (aseg <= 2999)
  323. aseg[cortical_label_mask] = aseg[cortical_label_mask] - 1000
  325. # Preserve Cortical Labels: The ones not touching:
  326. # "cortical regions touching each other across the hemispheres, are lateralized
  327. # while all others are com- bined thus reducing the total number of labels from 95
  328. # (DKT without corpus callosum segmentations which are added later) to 78 during network training."
  329. aseg[aseg_temp == 2014] = 2014
  330. aseg[aseg_temp == 2028] = 2028
  331. aseg[aseg_temp == 2012] = 2012
  332. aseg[aseg_temp == 2016] = 2016
  333. aseg[aseg_temp == 2002] = 2002
  334. aseg[aseg_temp == 2023] = 2023
  335. aseg[aseg_temp == 2017] = 2017
  336. aseg[aseg_temp == 2024] = 2024
  337. aseg[aseg_temp == 2010] = 2010
  338. aseg[aseg_temp == 2013] = 2013
  339. aseg[aseg_temp == 2025] = 2025
  340. aseg[aseg_temp == 2022] = 2022
  341. aseg[aseg_temp == 2021] = 2021
  342. aseg[aseg_temp == 2005] = 2005
  344. # Here 251, 1001, 1032, 1033 are added.
  345. labels = np.array([0, 2, 4, 5, 7, 8, 10, 11, 12, 13, 14,
  346. 15, 16, 17, 18, 24, 26, 28, 31, 41, 43, 44,
  347. 46, 47, 49, 50, 51, 52, 53, 54, 58, 60, 63,
  348. 77, 251, 1001, 1002, 1003, 1005, 1006, 1007, 1008, 1009, 1010, 1011,
  349. 1012, 1013, 1014, 1015, 1016, 1017, 1018, 1019, 1020, 1021, 1022,
  350. 1023, 1024, 1025, 1026, 1027, 1028, 1029, 1030, 1031, 1032, 1033, 1034, 1035,
  351. 2002, 2005, 2010, 2012, 2013, 2014, 2016, 2017, 2021, 2022, 2023,
  352. 2024, 2025, 2028])
  354. h, w, d = aseg.shape
  355. lut_aseg = np.zeros(max(labels) + 1, dtype='int')
  356. for idx, value in enumerate(labels):
  357. lut_aseg[value] = idx
  359. # Remap Label Classes - Perform LUT Mapping - Coronal, Axial
  361. mapped_aseg = lut_aseg.ravel()[aseg.ravel()]
  363. mapped_aseg = mapped_aseg.reshape((h, w, d))
  365. # Map Sagittal Labels
  366. aseg[aseg == 2] = 41
  367. aseg[aseg == 3] = 42
  368. aseg[aseg == 4] = 43
  369. aseg[aseg == 5] = 44
  370. aseg[aseg == 7] = 46
  371. aseg[aseg == 8] = 47
  372. aseg[aseg == 10] = 49
  373. aseg[aseg == 11] = 50
  374. aseg[aseg == 12] = 51
  375. aseg[aseg == 13] = 52
  376. aseg[aseg == 17] = 53
  377. aseg[aseg == 18] = 54
  378. aseg[aseg == 26] = 58
  379. aseg[aseg == 28] = 60
  380. aseg[aseg == 31] = 63
  382. cortical_label_mask = (aseg >= 2000) & (aseg <= 2999)
  383. aseg[cortical_label_mask] = aseg[cortical_label_mask] - 1000
  385. labels_sag = np.array([0, 14, 15, 16, 24, 41, 43, 44, 46, 47, 49,
  386. 50, 51, 52, 53, 54, 58, 60, 63, 77, 251, 1001, 1002,
  387. 1003, 1005, 1006, 1007, 1008, 1009, 1010, 1011, 1012, 1013, 1014,
  388. 1015, 1016, 1017, 1018, 1019, 1020, 1021, 1022, 1023, 1024, 1025,
  389. 1026, 1027, 1028, 1029, 1030, 1031, 1032, 1033, 1034, 1035])
  391. h, w, d = aseg.shape
  392. lut_aseg = np.zeros(max(labels_sag) + 1, dtype='int')
  393. for idx, value in enumerate(labels_sag):
  394. lut_aseg[value] = idx
  396. # Remap Label Classes - Perform LUT Mapping - Coronal, Axial
  398. mapped_aseg_sag = lut_aseg.ravel()[aseg.ravel()]
  400. mapped_aseg_sag = mapped_aseg_sag.reshape((h, w, d))
  402. return mapped_aseg, mapped_aseg_sag
  405. def map_wmparc2label(aseg):
  406. """
  407. Function to perform look-up table mapping of aparc.DKTatlas+aseg.mgz data to label space
  408. :param np.ndarray aseg: ground truth aparc+aseg
  409. :param None/np.ndarray aseg_nocc: ground truth aseg without corpus callosum segmentation
  410. :return:
  411. """
  412. aseg = aseg.astype(np.int16)
  413. aseg_temp = aseg.copy()
  415. aseg[aseg == 80] = 77 # Hypointensities Class
  416. aseg[aseg == 85] = 0 # Optic Chiasma to BKG
  417. aseg[aseg == 62] = 41 # Right Vessel to Right WM
  418. aseg[aseg == 30] = 2 # Left Vessel to Left MW
  419. aseg[aseg == 72] = 24 # 5th Ventricle to CSF
  421. # Fan: change wm-lh wm-rh to cerebral WM
  422. aseg[(aseg_temp >= 3000) & (aseg_temp < 3999)] = 2
  423. aseg[(aseg_temp >= 4000) & (aseg_temp < 4999)] = 41
  425. aseg[aseg_temp == 5001] = 2
  426. aseg[aseg_temp == 5002] = 41
  428. # Fan: combine CC to Corpus_Callosum
  429. aseg[(aseg >= 251) & (aseg <= 255)] = 192
  431. aseg[aseg == 3] = 0 # Map Remaining Cortical labels to background
  432. aseg[aseg == 42] = 0
  434. # If ctx-unknowns are not filled yet, do it now
  435. if np.any(np.in1d([1000, 2000], aseg.ravel())):
  436. aseg = fill_unknown_labels_per_hemi(aseg, 1000, 2000)
  437. aseg = fill_unknown_labels_per_hemi(aseg, 2000, 3000)
  439. aseg = fill_WMhyper_per_hemi(aseg) # Remove hyperinstensity
  441. # Fan: Make all right to be left:
  442. cortical_label_mask = (aseg >= 2000) & (aseg <= 2999)
  443. aseg[cortical_label_mask] = aseg[cortical_label_mask] - 1000
  445. # Preserve Cortical Labels: The ones not touching:
  446. # "cortical regions touching each other across the hemispheres, are lateralized
  447. # while all others are com- bined thus reducing the total number of labels from 95
  448. # (DKT without corpus callosum segmentations which are added later) to 78 during network training."
  449. aseg[aseg_temp == 2014] = 2014
  450. aseg[aseg_temp == 2028] = 2028
  451. aseg[aseg_temp == 2012] = 2012
  452. aseg[aseg_temp == 2016] = 2016
  453. aseg[aseg_temp == 2002] = 2002
  454. aseg[aseg_temp == 2023] = 2023
  455. aseg[aseg_temp == 2017] = 2017
  456. aseg[aseg_temp == 2024] = 2024
  457. aseg[aseg_temp == 2010] = 2010
  458. aseg[aseg_temp == 2013] = 2013
  459. aseg[aseg_temp == 2025] = 2025
  460. aseg[aseg_temp == 2022] = 2022
  461. aseg[aseg_temp == 2021] = 2021
  462. aseg[aseg_temp == 2005] = 2005
  464. # Here 192, 1001, 1032, 1033 are added.
  465. labels = np.array([0, 2, 4, 5, 7, 8, 10, 11, 12, 13, 14, 15,
  466. 16, 17, 18, 24, 26, 28, 31, 41, 43, 44, 46, 47,
  467. 49, 50, 51, 52, 53, 54, 58, 60, 63, 192,
  468. 1001, 1002, 1003, 1005, 1006, 1007, 1008, 1009, 1010, 1011, 1012,
  469. 1013, 1014, 1015, 1016, 1017, 1018, 1019, 1020, 1021, 1022, 1023,
  470. 1024, 1025, 1026, 1027, 1028, 1029, 1030, 1031, 1032, 1033, 1034, 1035,
  471. 2002, 2005, 2010, 2012, 2013, 2014, 2016, 2017, 2021, 2022, 2023,
  472. 2024, 2025, 2028])
  474. h, w, d = aseg.shape
  475. lut_aseg = np.zeros(max(labels) + 1, dtype='int')
  476. for idx, value in enumerate(labels):
  477. lut_aseg[value] = idx
  479. # Remap Label Classes - Perform LUT Mapping - Coronal, Axial
  481. mapped_aseg = lut_aseg.ravel()[aseg.ravel()]
  482. mapped_aseg = mapped_aseg.reshape((h, w, d))
  484. # Map Sagittal Labels
  485. aseg[aseg == 2] = 41
  486. aseg[aseg == 3] = 42
  487. aseg[aseg == 4] = 43
  488. aseg[aseg == 5] = 44
  489. aseg[aseg == 7] = 46
  490. aseg[aseg == 8] = 47
  491. aseg[aseg == 10] = 49
  492. aseg[aseg == 11] = 50
  493. aseg[aseg == 12] = 51
  494. aseg[aseg == 13] = 52
  495. aseg[aseg == 17] = 53
  496. aseg[aseg == 18] = 54
  497. aseg[aseg == 26] = 58
  498. aseg[aseg == 28] = 60
  499. aseg[aseg == 31] = 63
  501. cortical_label_mask = (aseg >= 2000) & (aseg <= 2999)
  502. aseg[cortical_label_mask] = aseg[cortical_label_mask] - 1000
  504. labels_sag = np.array([0, 14, 15, 16, 24, 41, 43, 44, 46, 47, 49,
  505. 50, 51, 52, 53, 54, 58, 60, 63, 192,
  506. 1001, 1002, 1003, 1005, 1006, 1007, 1008, 1009, 1010, 1011, 1012, 1013,
  507. 1014, 1015, 1016, 1017, 1018, 1019, 1020, 1021, 1022, 1023, 1024, 1025,
  508. 1026, 1027, 1028, 1029, 1030, 1031, 1032, 1033, 1034, 1035])
  510. h, w, d = aseg.shape
  511. lut_aseg = np.zeros(max(labels_sag) + 1, dtype='int')
  512. for idx, value in enumerate(labels_sag):
  513. lut_aseg[value] = idx
  515. # Remap Label Classes - Perform LUT Mapping - Coronal, Axial
  517. mapped_aseg_sag = lut_aseg.ravel()[aseg.ravel()]
  518. mapped_aseg_sag = mapped_aseg_sag.reshape((h, w, d))
  520. return mapped_aseg, mapped_aseg_sag
  522. def map_wmparc2gtseg(aseg):
  523. """
  524. Function to perform look-up table mapping of aparc.DKTatlas+aseg.mgz data to label space
  525. :param np.ndarray aseg: ground truth aparc+aseg
  526. :param None/np.ndarray aseg_nocc: ground truth aseg without corpus callosum segmentation
  527. :return:
  528. """
  529. aseg = aseg.astype(np.int16)
  530. aseg_temp = aseg.copy()
  532. aseg[aseg == 80] = 77 # Hypointensities Class
  533. aseg[aseg == 85] = 0 # Optic Chiasma to BKG
  534. aseg[aseg == 62] = 41 # Right Vessel to Right WM
  535. aseg[aseg == 30] = 2 # Left Vessel to Left MW
  536. aseg[aseg == 72] = 24 # 5th Ventricle to CSF
  538. # Fan: change wm-lh wm-rh to cerebral WM
  539. aseg[(aseg_temp >= 3000) & (aseg_temp < 3999)] = 2
  540. aseg[(aseg_temp >= 4000) & (aseg_temp < 4999)] = 41
  542. aseg[aseg_temp == 5001] = 2
  543. aseg[aseg_temp == 5002] = 41
  545. # Fan: combine CC to Corpus_Callosum
  546. aseg[(aseg >= 251) & (aseg <= 255)] = 192
  548. aseg[aseg == 3] = 0 # Map Remaining Cortical labels to background
  549. aseg[aseg == 42] = 0
  551. # If ctx-unknowns are not filled yet, do it now
  552. if np.any(np.in1d([1000, 2000], aseg.ravel())):
  553. aseg = fill_unknown_labels_per_hemi(aseg, 1000, 2000)
  554. aseg = fill_unknown_labels_per_hemi(aseg, 2000, 3000)
  556. aseg = fill_WMhyper_per_hemi(aseg) # Remove hyperinstensity
  558. return aseg
  560. def sagittal_coronal_remap_lookup(x):
  561. """
  562. Dictionary mapping to convert left labels to corresponding right labels for aseg
  563. :param int x: label to look up
  564. :return:
  565. """
  566. return {
  567. 2: 41,
  568. 3: 42,
  569. 4: 43,
  570. 5: 44,
  571. 7: 46,
  572. 8: 47,
  573. 10: 49,
  574. 11: 50,
  575. 12: 51,
  576. 13: 52,
  577. 17: 53,
  578. 18: 54,
  579. 26: 58,
  580. 28: 60,
  581. 31: 63,
  582. }[x]
  585. def map_prediction_sagittal2full(prediction_sag, num_classes=85):
  586. """
  587. Function to remap the prediction on the sagittal network to full label space used by coronal and axial networks
  588. (full aparc.DKTatlas+aseg.mgz)
  589. :param prediction_sag: sagittal prediction (labels)
  590. :param int num_classes: number of classes (96 for full classes, 79 for hemi split)
  591. :return: Remapped prediction
  592. """
  593. if num_classes == 96:
  594. idx_list = np.asarray([0, 5, 6, 7, 8, 9, 10, 11, 12, 13, 1, 2, 3, 14, 15, 4, 16,
  595. 17, 18, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19,
  596. 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36,
  597. 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 20, 21, 22,
  598. 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39,
  599. 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50], dtype=np.int16)
  601. else:
  602. # idx_list = np.asarray([0, 5, 6, 7, 8, 9, 10, 11, 12, 13, 1, 2, 3, 14, 15, 4, 16,
  603. # 17, 18, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19,
  604. # 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36,
  605. # 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 20, 22, 27,
  606. # 29, 30, 31, 33, 34, 38, 39, 40, 41, 42, 45], dtype=np.int16)
  608. labels = np.array([0, 2, 4, 5, 7, 8, 10, 11, 12, 13, 14,
  609. 15, 16, 17, 18, 24, 26, 28, 31, 41, 43, 44,
  610. 46, 47, 49, 50, 51, 52, 53, 54, 58, 60, 63,
  611. 192, 1001, 1002, 1003, 1005, 1006, 1007, 1008, 1009, 1010, 1011,
  612. 1012, 1013, 1014, 1015, 1016, 1017, 1018, 1019, 1020, 1021, 1022,
  613. 1023, 1024, 1025, 1026, 1027, 1028, 1029, 1030, 1031, 1032, 1033, 1034, 1035,
  614. 2002, 2005, 2010, 2012, 2013, 2014, 2016, 2017, 2021, 2022, 2023,
  615. 2024, 2025, 2028])
  617. labels_full_to_sag = np.array([0, 41, 43, 44, 46, 47, 49, 50, 51, 52, 14, 15, 16, 53, 54, 24, 58, 60, 63, 41, 43, 44, 46, 47, 49, 50, 51, 52, 53,
  618. 54, 58, 60, 63, 192, 1001, 1002, 1003, 1005, 1006, 1007, 1008, 1009, 1010, 1011, 1012, 1013, 1014, 1015, 1016,
  619. 1017, 1018, 1019, 1020, 1021, 1022, 1023, 1024, 1025, 1026, 1027, 1028, 1029, 1030, 1031, 1032, 1033, 1034, 1035,
  620. 1002, 1005, 1010, 1012, 1013, 1014, 1016, 1017, 1021, 1022, 1023, 1024, 1025, 1028])
  622. labels_sag = np.array([0, 14, 15, 16, 24, 41, 43, 44, 46, 47, 49,
  623. 50, 51, 52, 53, 54, 58, 60, 63, 192, 1001, 1002,
  624. 1003, 1005, 1006, 1007, 1008, 1009, 1010, 1011, 1012, 1013, 1014,
  625. 1015, 1016, 1017, 1018, 1019, 1020, 1021, 1022, 1023, 1024, 1025,
  626. 1026, 1027, 1028, 1029, 1030, 1031, 1032, 1033, 1034, 1035])
  628. idx = []
  629. for l in labels_full_to_sag:
  630. idx.append(np.where(labels_sag==l)[0][0])
  632. idx_list = np.array(idx)
  634. prediction_full = prediction_sag[:, idx_list, :, :]
  635. return prediction_full
  638. # Clean up and class separation
  639. def bbox_3d(img):
  640. """
  641. Function to extract the three-dimensional bounding box coordinates.
  642. :param np.ndarray img: mri image
  643. :return:
  644. """
  646. r = np.any(img, axis=(1, 2))
  647. c = np.any(img, axis=(0, 2))
  648. z = np.any(img, axis=(0, 1))
  650. rmin, rmax = np.where(r)[0][[0, -1]]
  651. cmin, cmax = np.where(c)[0][[0, -1]]
  652. zmin, zmax = np.where(z)[0][[0, -1]]
  654. return rmin, rmax, cmin, cmax, zmin, zmax
  657. def get_largest_cc(segmentation):
  658. """
  659. Function to find largest connected component of segmentation.
  660. :param np.ndarray segmentation: segmentation
  661. :return:
  662. """
  663. labels = label(segmentation, connectivity=3, background=0)
  665. bincount = np.bincount(labels.flat)
  666. background = np.argmax(bincount)
  667. bincount[background] = -1
  669. largest_cc = labels == np.argmax(bincount)
  671. return largest_cc
  674. # Class Operator for image loading (orig only)
  675. class OrigDataThickSlices(Dataset):
  676. """
  677. Class to load a given image and segmentation and prepare it
  678. for network training.
  679. """
  680. def __init__(self, img_filename, orig, plane='Axial', slice_thickness=3, transforms=None):
  682. try:
  683. self.img_filename = img_filename
  684. self.plane = plane
  685. self.slice_thickness = slice_thickness
  687. # Transform Data as needed
  688. if plane == 'Sagittal':
  689. orig = transform_sagittal(orig)
  690. print('Loading Sagittal')
  692. elif plane == 'Axial':
  693. orig = transform_axial(orig)
  694. print('Loading Axial')
  696. else:
  697. print('Loading Coronal.')
  699. # Create Thick Slices
  700. orig_thick = get_thick_slices(orig, self.slice_thickness)
  702. # Make 4D
  703. orig_thick = np.transpose(orig_thick, (2, 0, 1, 3))
  704. self.images = orig_thick
  706. self.count = self.images.shape[0]
  708. self.transforms = transforms
  710. print("Successfully loaded Image from {}".format(img_filename))
  712. except Exception as e:
  713. print("Loading failed. {}".format(e))
  715. def __getitem__(self, index):
  717. img = self.images[index]
  719. if self.transforms is not None:
  720. img = self.transforms(img)
  722. return {'image': img}
  724. def __len__(self):
  725. return self.count
  727. # Class Operator for image loading (orig only)
  728. class OrigDataThickSlices_Fused_Input(Dataset):
  729. """
  730. Class to load a given image and segmentation and prepare it
  731. for network training.
  732. """
  733. def __init__(self, img_filename, orig, plane='Axial', slice_thickness=3, transforms=None):
  735. try:
  736. self.img_filename = img_filename
  737. self.plane = plane
  738. self.slice_thickness = slice_thickness
  740. orig_thick_list = []
  741. for idx, orig_ in enumerate(orig):
  742. # Transform Data as needed
  743. if plane == 'Sagittal':
  744. orig_ = transform_sagittal(orig_)
  745. print('Loading Sagittal %d' % idx)
  747. elif plane == 'Axial':
  748. orig_ = transform_axial(orig_)
  749. print('Loading Axial %d' % idx)
  751. else:
  752. print('Loading Coronal %d' % idx)
  754. # Create Thick Slices
  755. orig_thick = get_thick_slices(orig_, self.slice_thickness)
  757. # Make 4D
  758. orig_thick = np.transpose(orig_thick, (2, 0, 1, 3))
  759. orig_thick_list.append(orig_thick)
  761. self.images = np.concatenate(orig_thick_list, axis=3)
  763. self.count = self.images.shape[0]
  765. self.transforms = transforms
  767. print("Successfully loaded Image from {}".format(img_filename))
  769. except Exception as e:
  770. print("Loading failed. {}".format(e))
  772. def __getitem__(self, index):
  774. img = self.images[index]
  776. if self.transforms is not None:
  777. img = self.transforms(img)
  779. return {'image': img}
  781. def __len__(self):
  782. return self.count
  784. ##
  785. # Dataset loading (for training)
  786. ##
  788. # Operator to load hdf5-file for training
  789. class AsegDatasetWithAugmentation(Dataset):
  790. """
  791. Class for loading aseg file with augmentations (transforms)
  792. """
  793. def __init__(self, params, transforms=None):
  795. # Load the h5 file and save it to the dataset
  796. try:
  797. self.params = params
  799. # Open file in reading mode
  800. with h5py.File(self.params['dataset_name'], "r") as hf:
  801. self.images = np.array(hf.get('orig_dataset'))
  802. self.labels = np.array(hf.get('aseg_dataset'))
  803. self.weights = np.array(hf.get('weight_dataset'))
  804. self.subjects = np.array(hf.get("subject"))
  806. self.count = self.images.shape[0]
  807. self.transforms = transforms
  809. print("Successfully loaded {} with plane: {}".format(params["dataset_name"], params["plane"]))
  811. except Exception as e:
  812. print("Loading failed: {}".format(e))
  814. def get_subject_names(self):
  815. return self.subjects
  817. def __getitem__(self, index):
  819. img = self.images[index]
  820. label = self.labels[index]
  821. weight = self.weights[index]
  823. if self.transforms is not None:
  824. tx_sample = self.transforms({'img': img, 'label': label, 'weight': weight})
  825. img = tx_sample['img']
  826. label = tx_sample['label']
  827. weight = tx_sample['weight']
  829. return {'image': img, 'label': label, 'weight': weight}
  831. def __len__(self):
  832. return self.count
  834. # Operator to load hdf5-file for training
  835. class AsegDatasetWithAugmentation_Fused_Input(Dataset):
  836. """
  837. Class for loading aseg file with augmentations (transforms)
  838. """
  839. def __init__(self, params, transforms=None):
  841. # Load the h5 file and save it to the dataset
  842. try:
  843. self.params = params
  845. self.images = []
  846. # Open file in reading mode
  847. for idx in range(len(self.params['dataset_name'])):
  849. with h5py.File(self.params['dataset_name'][idx], "r") as hf:
  850. self.images.append(np.array(hf.get('orig_dataset')))
  851. if idx == 0:
  852. self.labels = np.array(hf.get('aseg_dataset'))
  853. self.weights = np.array(hf.get('weight_dataset'))
  854. self.subjects = np.array(hf.get("subject"))
  856. self.images = np.concatenate(self.images, axis=3)
  858. self.count = self.images.shape[0]
  859. self.transforms = transforms
  861. print("Successfully loaded {} with plane: {}".format(params["dataset_name"], params["plane"]))
  863. except Exception as e:
  864. print("Loading failed: {}".format(e))
  866. def get_subject_names(self):
  867. return self.subjects
  869. def __getitem__(self, index):
  871. img = self.images[index]
  872. label = self.labels[index]
  873. weight = self.weights[index]
  875. if self.transforms is not None:
  876. tx_sample = self.transforms({'img': img, 'label': label, 'weight': weight})
  877. img = tx_sample['img']
  878. label = tx_sample['label']
  879. weight = tx_sample['weight']
  881. return {'image': img, 'label': label, 'weight': weight}
  883. def __len__(self):
  884. return self.count
  887. class AsegDatasetWithAugmentation_Slice(Dataset):
  888. """
  889. Class for loading aseg file with augmentations (transforms)
  890. """
  891. def __init__(self, params, transforms=None):
  893. # Load the h5 file and save it to the dataset
  894. try:
  895. self.params = params
  897. self.hdf5_files = sorted(glob.glob(os.path.join(params['dataset_name'], '*hdf5')))
  898. self.count = len(self.hdf5_files)
  899. self.transforms = transforms
  901. print("Successfully loaded {} with plane: {}".format(params["dataset_name"], params["plane"]))
  903. except Exception as e:
  904. print("Loading failed: {}".format(e))
  906. def get_subject_names(self):
  907. return self.subjects
  909. def __getitem__(self, index):
  911. hdf5 = self.hdf5_files[index]
  912. with h5py.File(hdf5, "r") as hf:
  913. img = np.array(hf.get('orig_dataset')).squeeze()
  914. label = np.array(hf.get('aseg_dataset')).squeeze()
  915. weight = np.array(hf.get('weight_dataset')).squeeze()
  917. if self.transforms is not None:
  918. tx_sample = self.transforms({'img': img, 'label': label, 'weight': weight})
  919. img = tx_sample['img']
  920. label = tx_sample['label']
  921. weight = tx_sample['weight']
  923. return {'image': img, 'label': label, 'weight': weight}
  925. def __len__(self):
  926. return self.count
  928. class AsegDatasetWithAugmentation_Slice_Fused_Input(Dataset):
  929. """
  930. Class for loading aseg file with augmentations (transforms)
  931. """
  932. def __init__(self, params, transforms=None):
  934. # Load the h5 file and save it to the dataset
  935. try:
  936. self.params = params
  937. self.params['dataset_name'] = sorted(self.params['dataset_name'])
  938. self.hdf5_files = []
  939. tmp_len = []
  940. for dataset in self.params['dataset_name']:
  941. filelist = sorted(glob.glob(os.path.join(dataset, '*hdf5')))
  942. self.hdf5_files.append(filelist)
  943. tmp_len.append(len(filelist))
  945. if np.unique(tmp_len).shape[0] != 1:
  946. print("Error: Input data should have the same number of hdf5 files!")
  947. exit()
  949. self.hdf5_files = np.array(self.hdf5_files)
  950. self.count = np.unique(tmp_len)[0]
  951. self.transforms = transforms
  953. print("Successfully loaded {} with plane: {}".format(params["dataset_name"], params["plane"]))
  955. except Exception as e:
  956. print("Loading failed: {}".format(e))
  958. def get_subject_names(self):
  959. return self.subjects
  961. def __getitem__(self, index):
  963. hdf5s = self.hdf5_files[:, index]
  964. if np.unique([os.path.split(n)[1] for n in hdf5s]).shape[0] != 1:
  965. print("Error: Must have the same file name!")
  966. print(hdf5s)
  967. exit()
  969. img = []
  970. for idx, hdf5 in enumerate(hdf5s):
  971. with h5py.File(hdf5, "r") as hf:
  972. img.append(np.array(hf.get('orig_dataset')).squeeze())
  973. if idx == 0:
  974. label = np.array(hf.get('aseg_dataset')).squeeze()
  975. weight = np.array(hf.get('weight_dataset')).squeeze()
  976. img = np.concatenate(img, axis=2)
  978. if self.transforms is not None:
  979. tx_sample = self.transforms({'img': img, 'label': label, 'weight': weight})
  980. img = tx_sample['img']
  981. label = tx_sample['label']
  982. weight = tx_sample['weight']
  984. return {'image': img, 'label': label, 'weight': weight}
  986. def __len__(self):
  987. return self.count

load_neuroimaging_data.py at commit 257ec7c, under Unlicense · at the source

Overview

Authors: Christian Ewert1, David Kügler1, Martin Reuter1,2,3
ORCID iDs: Christian Ewert, David Kügler, Martin Reuter
  1. German Center for Neurodegenerative Diseases (DZNE), Bonn, Germany
  2. A. A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Boston, MA, United States
  3. Department of Radiology, Harvard Medical School, Boston, MA, United States
Institutions: German Center for Neurodegenerative Diseases (Germany); Harvard University (United States); Massachusetts General Hospital (United States); Athinoula A. Martinos Center for Biomedical Imaging (United States)
Journal: Imaging neuroscience (Cambridge, Mass.), volume 4, article IMAG.a.1183
Dates: received 1 July 2025; accepted 13 February 2026; published online 2 June 2026
Type: Research article · Language: English
License: CC BY
Identifiers: DOI 10.1162/imag.a.1183 · PMID 42245204 · PMCID PMC13231284 · OpenAlex W7134815812
Open access: diamond, a free copy (OpenAlex)
Status: code verified
Categories: structural MRI / diffusion (modality), methods / tools (subfield)
Methods: Connectivity, Statistics, Machine learning, fMRI & imaging
Keywords: deep learning, geometric deep learning, segmentation, diffusion MRI, q-space
Topic: Advanced Neuroimaging Techniques and Applications (Radiology, Nuclear Medicine and Imaging, Medicine), according to OpenAlex
Funding: Federal Ministry of Research, Technology and Space of Germany (BMFTR 031L0206); Chan Zuckerberg Initiative (EOSS5 2022-252594); Helmholtz Foundation Model Initiative (The Human Radiome Project); U.S. Department of Health & Human Services | NIH | National Institute on Aging (U.S. National Institute on Aging) (R01-AG064027); HHS | NIH | National Institute of Mental Health (NIMH) (R01-MH131586, R01-MH130899); Deutsches Zentrum für Neurodegenerative Erkrankungen (German Center for Neurodegenerative Diseases)
Citations: not cited yet (Europe PMC); 61 references in the paper

Abstract

Segmenting anatomical regions is a crucial step in many diffusion-weighted MRI (dMRI) workflows, such as region-of-interest analysis or anatomically-constrained tractography, which enable in vivo studies of brain microstructure and connectivity. However, convolutional neural networks (CNNs)—the foundation of most state-of-the-art segmentation models—require structured inputs with a fixed number of channels. This makes them ill-suited for dMRI, where acquisition protocols vary widely in q-space sampling—the number of measurements as well as their directions (b-vectors) and weightings (b-values)—resulting in unstructured data with inconsistent dimensionality across studies. As a consequence, the applicability of CNN-based methods is generally limited to the dataset on which they were trained. To address this, existing methods like DeepAnat and DDParcel rely on diffusion model fits, such as the diffusion tensor, to convert raw data into structured representations compatible with CNNs. While this enables broader applicability, it introduces lossy compression that can degrade performance. In this work, we propose a novel method that combines the geometric deep learning-based reconstruction framework DISCUS with the segmentation network VINN to directly map unstructured dMRI data to anatomical segmentations. Our segmentation approach is the first to achieve robust generalization across heterogeneous acquisition schemes using a single neural network without requiring diffusion model fits. Our approach generates the segmentation in minutes, whereas DeepAnat relies on the external FreeSurfer software, which runs for several hours. Additionally, we demonstrate generally superior segmentation performance of our approach across multiple datasets and acquisition settings with respect to DeepAnat, DDParcel, and SynthSeg.

Reproduced under the paper's license (CC BY), from the paper cited above.

Repositories

Its files are read in the Code ↔ Paper reader above, with 11 matches between paragraphs and lines of code.

liziyu0929/DeepAnat

License: MIT
State: the link answers, verified on 27 September 2026
Evidence: files inventoried
Commit: 365f8e4529faa4cce3470450e6aaef4f395b51e4, 9 February 2023
Languages: Python (6), Jupyter (1)
Size: 13 files, 7 scripts
Software Heritage: not archived
Found in: the end of the paper
Holds: README, license file, 1 notebook
Not found: CITATION.cff, environment file, tests, continuous integration, documentation
Tools: Keras (7 files), TensorFlow (7 files), NiBabel (5 files), NumPy (5 files), Matplotlib (4 files), SciPy (4 files)
Availability: 1 check, the latest on 27 September 2026: the link answers
  • 27 September 2026: the link answers
9 files
At the source: github.com/liziyu0929/DeepAnat

zhangfanmark/DDParcel

License: Unlicense
State: the link answers, verified on 27 September 2026
Evidence: files inventoried
Commit: 257ec7c2a02209f86365dc5c0b39755458a3f776, 9 December 2024
Languages: Python (11), Shell (2)
Size: 16 files, 13 scripts
Software Heritage: not archived
Found in: the end of the paper
Holds: README, license file
Not found: CITATION.cff, environment file, tests, continuous integration, documentation
Tools: PyTorch (7 files), NumPy (6 files), NiBabel (4 files), scikit-image (4 files), SciPy (4 files), h5py (1 file), Matplotlib (1 file)
Availability: 1 check, the latest on 27 September 2026: the link answers
  • 27 September 2026: the link answers
15 files
At the source: github.com/zhangfanmark/DDParcel

Deep-MI/DISCUSseg

License: Apache-2.0
State: the link answers, verified on 27 September 2026
Evidence: files inventoried
Commit: c45ed3b5ba88c09c163074d5e08b3d6c521bdb9c, 1 June 2026
Languages: Python (37)
Size: 44 files, 37 scripts
Software Heritage: not archived
Found in: “Data and Code Availability”
Holds: README, license file
Not found: CITATION.cff, environment file, tests, continuous integration, documentation
Tools: PyTorch (22 files), NumPy (16 files), NiBabel (5 files), Matplotlib (3 files), scikit-image (3 files), SciPy (2 files), pandas (1 file)
Availability: 1 check, the latest on 27 September 2026: the link answers
  • 27 September 2026: the link answers
39 files
At the source: github.com/Deep-MI/DISCUSseg

The paper's code and data availability statement is in the Data section.

Tracing map

Proposed by the machine: these links were found in the paper and verified at the source, without human review. The map will receive a Zenodo DOI once one of the paper's authors has validated it with their ORCID.

What the map holds:

  • 3 repositories of the authors' code, each at its verified commit, with its license and how the link was found in the paper;
  • 57 scripts, each with its path and the digest of its content;
  • 11 matches between paragraphs of the paper and lines of the code (method lexical-v1);
  • neither the text of the paper nor the code itself.

Its JSON (tracing-map.json) is deposited on Zenodo with its DOI once the map is validated.

Data

No dataset and no data link were found in the paper.

Data and Code Availability

In this paper, we use MRI data of two public datasets, the Young Adult dataset of the Human Connectome Project (HCP), available at https://balsa.wustl.edu, and the dataset of the Alzheimer’s Disease Neuroimaging Initiative (ADNI), available at https://ida.loni.usc.edu. In addition, we use MRI data of the Rhineland Study which is not publicly available due to data protection regulations. Access can be provided to scientists in accordance with the Rhineland Study’s Data Use and Access Policy. Requests to access the data should be directed to Monique M.B. Breteler at . The source code for our method can be found on the website https://github.com/Deep-MI/DISCUSseg.

Reproduced under the paper's license (CC BY), from the paper cited above.

Versions

The history of this record: each version stored by the harvester or made by a correction of its authors or of the maintainers of its code, and what changed in its facts. The texts of the paper (its abstract, its availability statements) are not part of it; versions that changed only those are not listed.

Version 1, 27 September 2026: the first record

Recorded: type, language, journal, volume, pages, dates, 3 authors, 5 keywords, 6 funders, 59 references.

Cite

This paper

Ewert, C., Kügler, D., & Reuter, M. (2026). Learning-based segmentation of diffusion-weighted MR images with arbitrary &lt;i&gt;q&lt;/i&gt;-space samplings. Imaging neuroscience (Cambridge, Mass.), 4, IMAG.a.1183. https://doi.org/10.1162/imag.a.1183

BibTeX

@article{ewert2026learning,
author = {Ewert, Christian and Kügler, David and Reuter, Martin},
title = {{Learning-based segmentation of diffusion-weighted MR images with arbitrary \&lt;i\&gt;q\&lt;/i\&gt;-space samplings}},
journal = {Imaging neuroscience (Cambridge, Mass.)},
year = {2026},
month = jun,
volume = {4},
pages = {IMAG.a.1183},
publisher = {MIT Press},
issn = {2837-6056},
doi = {10.1162/imag.a.1183},
url = {https://doi.org/10.1162/imag.a.1183},
pmid = {42245204},
pmcid = {PMC13231284}
}

RIS

TY - JOUR
AU - Ewert, Christian
AU - Kügler, David
AU - Reuter, Martin
TI - Learning-based segmentation of diffusion-weighted MR images with arbitrary &lt;i&gt;q&lt;/i&gt;-space samplings
T2 - Imaging neuroscience (Cambridge, Mass.)
J2 - Imaging Neurosci (Camb)
PY - 2026
DA - 2026/06/02
VL - 4
SP - IMAG.a.1183
SN - 2837-6056
PB - MIT Press
DO - 10.1162/imag.a.1183
UR - https://doi.org/10.1162/imag.a.1183
LA - en
ER -

CSL-JSON

{
"id": "10.1162/imag.a.1183",
"type": "article-journal",
"title": "Learning-based segmentation of diffusion-weighted MR images with arbitrary &lt;i&gt;q&lt;/i&gt;-space samplings",
"container-title": "Imaging neuroscience (Cambridge, Mass.)",
"author": [
{
"family": "Ewert",
"given": "Christian"
},
{
"family": "Kügler",
"given": "David"
},
{
"family": "Reuter",
"given": "Martin"
}
],
"container-title-short": "Imaging Neurosci (Camb)",
"volume": "4",
"page": "IMAG.a.1183",
"DOI": "10.1162/imag.a.1183",
"PMID": "42245204",
"PMCID": "PMC13231284",
"ISSN": "2837-6056",
"publisher": "MIT Press",
"URL": "https://doi.org/10.1162/imag.a.1183",
"language": "en",
"issued": {
"date-parts": [
[
2026,
6,
2
]
]
}
}

The tracing map gets a citation of its own once an author has validated it and it has a DOI.

Similar papers

The papers with a page that share the most with this one: the tools found in their code, their categories, datasets, cited references and authors, the rarest counting most.

[1] doi:10.1038/s41598-026-54446-8 [code]
Deep learning-based Desikan-Killiany parcellation of the brain using diffusion MRI.
Journal: Scientific reports
In common: scikit-image, h5py, NiBabel, 3 other tools, methods / tools, structural MRI / diffusion, 12 references
[2] doi:10.1002/advs.76596 [code]
DDSurfer: A Weakly-Supervised Dual-Stream Deep Learning Framework for Cortical Surface Reconstruction From Diffusion MRI.
Journal: Advanced science (Weinheim, Baden-Wurttemberg, Germany)
In common: NiBabel, PyTorch, pandas, 2 other tools, methods / tools, structural MRI / diffusion, 10 references
[3] doi:10.1002/nbm.70353 [code]
Automated Surface-Based Segmentation of Deep Gray Matter Regions Based on Diffusion Tensor Images Reveals Unique Age Trajectories Over the Healthy Lifespan.
Journal: NMR in biomedicine
In common: scikit-image, NiBabel, SciPy, 1 other tool, structural MRI / diffusion, 9 references
[4] doi:10.1038/s41597-026-07184-5 [code]
A Multiple Sclerosis MRI Dataset with Tri-Mask Annotations for Lesion Segmentation.
Journal: Scientific data
In common: Keras, TensorFlow, scikit-image, 5 other tools, methods / tools, structural MRI / diffusion, 3 references
[5] doi:10.1162/imag.a.1198 [code]
MEPrep: A robust pipeline for multi-echo fMRI denoising and preprocessing.
Journal: Imaging neuroscience (Cambridge, Mass.)
In common: scikit-image, h5py, NiBabel, 4 other tools, methods / tools, 5 references
[6] doi:10.1038/s41598-026-55397-w [code]
Fast surface reconstruction of human brain MRI: benchmarking deep-learning based morphometry tools.
Journal: Scientific reports
In common: scikit-image, NiBabel, PyTorch, 3 other tools, methods / tools, structural MRI / diffusion, 6 references
[7] doi:10.1162/imag.a.1366 [code]
MICAFlow: Fast and robust MRI preprocessing bridging research neuroimaging and clinical practice.
Journal: Imaging neuroscience (Cambridge, Mass.)
In common: Keras, TensorFlow, NiBabel, 4 other tools, methods / tools, structural MRI / diffusion, 4 references
[8] doi:10.1038/s41586-026-10631-3 [code]
A prognostic human brain network for diffuse midline glioma.
Journal: Nature
In common: Keras, TensorFlow, h5py, 5 other tools, 4 references
[9] doi:10.1002/alz.71530 [code]
Differential associations of plasma biomarkers with Alzheimer's disease and small vessel disease: A multimodal imaging study.
Journal: Alzheimer's & dementia : the journal of the Alzheimer's Association
In common: NiBabel, PyTorch, pandas, 3 other tools, structural MRI / diffusion, 6 references
[10] doi:10.1080/07853890.2026.2685416 [code]
Pulmonary and cerebral damage in COVID-19 survivors: is there any association?
Journal: Annals of medicine
In common: Keras, TensorFlow, scikit-image, 7 other tools, structural MRI / diffusion

Contribute

The authors of this paper can claim it, correct its record and validate its tracing map, and the maintainers of its code (its owner, or a public member of its organization) correct what it says of their repository; anyone signed in can ask for its removal. Every request goes to OSCR's own machine, which answers it; your account page follows them.

Sign in with ORCID to claim this paper as one of its authors, correct its record or validate its tracing map: when the paper's metadata lists your ORCID iD, you are recognized at once. Maintainers of its code: sign in with GitHub, then claim the repository on your account page.

Request its removal

To ask OSCR to remove this record, the copies of its authors' scripts or its tracing map, use the removal request page: signed in, you say who you are, what to remove and why, then review and confirm the request. Published rules decide every request (how).

Discussion, reproductions, activity

Discussion: questions and error reports about this paper and its code, from signed-in readers and its authors. It opens with sign-in.

Reproductions: reports from readers who ran the authors' code: what they reproduced, with which environment, commit and data. It opens with sign-in.

Activity: what happens around this paper: new versions of its record, its map's validation, discussions and reproductions. It opens with sign-in.