OSCR

Data-centric strategies enable cross-center generalization of convolutional neural networks for CT brain extraction.

Code ↔ Paper

The paper beside its authors' code: matches between them have not been computed for this paper yet.

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 · 562 lines · 24 KB · no license

  1. """
  2. AugGen
  3. ======
  4. This module implements the AugmentationGenerator class which can be used to
  5. create augmentations of images and their segmentations by performing
  6. affine transformations.
  7. """
  8. import numpy as np
  9. from scipy import ndimage
  10. import nibabel.orientations as orx
  13. class AugmentationGenerator(object):
  14. """Initializes an augmentation generator object.
  16. ### Args
  17. `rotation_z` (int): Max angle of rotation around z-axis in degrees.
  18. Defaults to 15.
  19. `rotation_x` (int): Max angle of rotation around x-axis in degrees.
  20. Defaults to 2.
  21. `rotation_y` (int): Max angle of rotation around y-axis in degrees.
  22. Defaults to 2.
  23. `translation_xy` (int): Max number of pixels to translate in the
  24. xy plane. Defaults to 15.
  25. `translation_z` (int): Max number of pixels to translate in the
  26. z axis. Defaults to 0.
  27. `scale_xy` (float): Scaling factor is 1 +/- `scale_delta_max`,
  28. Defaults to 0.1.
  29. `scale_z` (int): Scaling factor is 1 +/- `scale_delta_max`,
  30. Defaults to 0.
  31. `thickness` (array-like): Slice thickness to pick from. Defaults to (1,)
  32. `flip_h` (bool): Determines whether horizontal flipping can occur.
  33. Defaults to True.
  34. `flip_v` (bool): Determines whether vertical flipping can occur.
  35. Defaults to False.
  36. `filling_mode` (str): mode passed to ndimage.affine_transform function,
  37. ('constant', 'nearest', 'reflect', 'mirror', 'wrap). Defaults to
  38. 'constant
  39. `padding_value` (float): Determines the value used for filling if
  40. filling mode is 'constant
  41. `bone_edges` (bool): Determines whether mask pixels are adjusted
  42. to account for bone edges. Defaults to False.
  43. """
  44. def __init__(self, rotation_z=0, rotation_x=0, rotation_y=0,
  45. translation_xy=0, translation_z=0, scale_xy=0,
  46. scale_z=0, thickness=(0,), flip_h=False, flip_v=False,
  47. filling_mode='constant', padding_value=0.0, bone_edges=False, mask_PV=0):
  48. self.rotation_z = rotation_z
  49. self.rotation_x = rotation_x
  50. self.rotation_y = rotation_y
  51. self.translation_xy = translation_xy
  52. self.translation_z = translation_z
  53. self.scale_xy = scale_xy
  54. self.scale_z = scale_z
  55. self.thickness = thickness
  56. self.flip_h = bool(flip_h)
  57. self.flip_v = bool(flip_v)
  58. self.filling_mode = filling_mode
  59. self.padding_value = padding_value
  60. self.mask_PV = mask_PV
  61. self.bone_edges = bone_edges
  63. def __repr__(self):
  64. string = 'AugmentationGenerator('
  65. string += 'rotation_z: ' + str(self.rotation_z) + ', '
  66. string += 'rotation_x: ' + str(self.rotation_x) + ', '
  67. string += 'rotation_y: ' + str(self.rotation_y) + ', '
  68. string += 'translation_xy: ' + str(self.translation_xy) + ', '
  69. string += 'translation_z: ' + str(self.translation_z) + ', '
  70. string += 'scale_xy: ' + str(self.scale_xy) + ', '
  71. string += 'scale_z: ' + str(self.scale_z) + ', '
  72. string += 'thickness: ' + str(self.thickness) + ', '
  73. string += 'flip_h: ' + str(self.flip_h) + ', '
  74. string += 'flip_v: ' + str(self.flip_v) + ')'
  75. string += 'filling_mode: ' + str(self.filling_mode) + ')'
  76. string += 'padding_value: ' + str(self.padding_value) + ')'
  77. string += 'bone_edges: ' + str(self.bone_edges) + ')'
  79. return string
  81. def __str__(self):
  82. string = 'ImageAugmentationGenerator parameters:\n'
  83. string += ' rotation_z: ' + str(self.rotation_z) + '\n'
  84. string += ' rotation_x: ' + str(self.rotation_x) + '\n'
  85. string += ' rotation_y: ' + str(self.rotation_y) + '\n'
  86. string += ' translation_xy: ' + str(self.translation_xy) + '\n'
  87. string += ' translation_z: ' + str(self.translation_z) + '\n'
  88. string += ' scale_xy: ' + str(self.scale_xy) + '\n'
  89. string += ' scale_z: ' + str(self.scale_z) + '\n'
  90. string += ' thickness: ' + str(self.thickness) + '\n'
  91. string += ' flip_h: ' + str(self.flip_h) + '\n'
  92. string += ' flip_v: ' + str(self.flip_v) + '\n'
  93. string += ' filling_mode: ' + str(self.filling_mode) + '\n'
  94. string += ' padding_value: ' + str(self.padding_value) + '\n'
  95. string += ' bone_edges: ' + str(self.bone_edges) + '\n'
  97. return string
  99. def _create_comp_affine(self, affine, offset):
  100. """Returns a composite matrix from the affine and offset."""
  101. if affine.shape != (2, 2) and affine.shape != (3, 3):
  102. raise ValueError('Affine must be a 2x2 or 3x3 matrix.')
  103. if affine.shape[0] != offset.shape[0]:
  104. raise ValueError('Affine and offset have incompatible dimensions.')
  106. affine = np.concatenate((affine, offset.reshape(-1, 1)), axis=1)
  108. if affine.shape == (3, 4):
  109. affine = np.concatenate((affine, np.array([[0, 0, 0, 1]])), axis=0)
  110. else:
  111. affine = np.concatenate((affine, np.array([[0, 0, 1]])), axis=0)
  113. return affine
  115. def generate(self, src, mask, n, affine, return_orig=True, verbose=0):
  116. """Generate `n` augmentations of the `src` and `mask`.
  118. ### Args
  119. `src` (three dimensional ndarray): Image matrix.
  120. `mask` (three dimensional ndarray): Binary mask matrix.
  121. `n` (int): Number of augmentations.
  122. `return_orig` (bool): Determines whether original src and mask are
  123. returned. Defaults to True.
  124. `verbose` (int): Determines if augmentation parameters are printed
  125. to screen. Defaults to 0.
  127. ### Returns:
  128. Tuple of two matrices, first matrix contains stacked `src`
  129. augmentations while the second contains the stacked `mask`
  130. augmentations.
  131. """
  132. # Check if src and mask are compatible
  133. if src.ndim != mask.ndim:
  134. raise ValueError('Src and mask dimensions don\'t match.')
  136. # If the passed src is 2 dimensional, call the generate2d method instead
  137. if src.ndim == 2:
  138. return self._generate2d(src, mask, n, return_orig, verbose)
  140. # (x1,y1,z1) = orx.aff2axcodes(affine)
  141. # ornt = orx.axcodes2ornt((x1,y1,z1))
  142. # refOrnt = orx.axcodes2ornt(('R','A','S'))
  143. # newOrnt1 = orx.ornt_transform(ornt,refOrnt)
  145. # (x2,y2,z2) = orx.aff2axcodes(affine)
  146. # ornt = orx.axcodes2ornt((x2,y2,z2))
  147. # refOrnt = orx.axcodes2ornt(('R','A','S'))
  148. # newOrnt2 = orx.ornt_transform(ornt,refOrnt)
  150. # src = orx.apply_orientation(src,newOrnt1)
  151. # mask = orx.apply_orientation(mask,newOrnt1)
  153. # src = np.fliplr(np.rot90(src,1))
  154. # mask = np.fliplr(np.rot90(mask,1))
  157. # Get parameters from self
  158. rotation_z = self.rotation_z
  159. rotation_x = self.rotation_x
  160. rotation_y = self.rotation_y
  161. translation_xy = self.translation_xy
  162. translation_z = self.translation_z
  163. scale_xy = self.scale_xy
  164. scale_z = self.scale_z
  165. thickness=self.thickness
  166. flip_h = self.flip_h
  167. flip_v = self.flip_v
  168. filling_mode = self.filling_mode
  169. padding_value = self.padding_value
  170. bone_edges = self.bone_edges
  171. mask_PV = self.mask_PV
  172. # Add original images and mask to output stack if warranted
  173. if return_orig:
  174. output_img_stack = src.copy()
  175. output_mask_stack = mask.copy().astype(float)
  177. if verbose:
  178. print(self)
  180. for iteration in range(n):
  181. # Create base affine
  182. affine_w_offset = np.identity(4)
  184. #----------------------------------------------------#
  185. # Rotate around z-axis (rotation in the axial plane) #
  186. #----------------------------------------------------#
  187. if rotation_z:
  188. # Choose an angle at random and convert it to radians
  189. rot_z_angle = np.random.randint(-rotation_z, rotation_z + 1)
  190. rot_z_angle = np.deg2rad(rot_z_angle)
  192. affine = \
  193. np.array([[np.cos(rot_z_angle), -np.sin(rot_z_angle), 0],
  194. [np.sin(rot_z_angle), np.cos(rot_z_angle), 0],
  195. [0, 0, 1]])
  197. # Calculate center offset
  198. center_input = 0.5 * np.array(src.shape)
  199. center_output = center_input.dot(affine)
  200. offset = center_input - center_output
  202. # Calculate new composite affine
  203. new_affine_w_offset = self._create_comp_affine(affine, offset)
  204. affine_w_offset = np.dot(affine_w_offset, new_affine_w_offset)
  206. #-------------------------------------------------------#
  207. # Rotate around x-axis (rotation in the sagittal plane) #
  208. #-------------------------------------------------------#
  209. if rotation_x:
  210. # Choose an angle at random and convert it to radians
  211. rot_x_angle = np.random.randint(-rotation_x, rotation_x + 1)
  212. rot_x_angle = np.deg2rad(rot_x_angle)
  214. affine = \
  215. np.array([[1, 0, 0],
  216. [0, np.cos(rot_x_angle), -np.sin(rot_x_angle)],
  217. [0, np.sin(rot_x_angle), np.cos(rot_x_angle)]])
  219. # Calculate center offset
  220. center_input = 0.5 * np.array(src.shape)
  221. center_output = center_input.dot(affine)
  222. offset = center_input - center_output
  224. # Calculate new composite affine
  225. new_affine_w_offset = self._create_comp_affine(affine, offset)
  226. affine_w_offset = np.dot(affine_w_offset, new_affine_w_offset)
  228. #------------------------------------------------------#
  229. # Rotate around y-axis (rotation in the coronal plane) #
  230. #------------------------------------------------------#
  231. if rotation_y:
  232. # Choose an angle at random and convert it to radians
  233. rot_y_angle = np.random.randint(-rotation_y, rotation_y + 1)
  234. rot_y_angle = np.deg2rad(rot_y_angle)
  236. affine = \
  237. np.array([[np.cos(rot_y_angle), 0, np.sin(rot_y_angle)],
  238. [0, 1, 0],
  239. [-np.sin(rot_y_angle), 0, np.cos(rot_y_angle)]])
  241. # Calculate center offset
  242. center_input = 0.5 * np.array(src.shape)
  243. center_output = center_input.dot(affine)
  244. offset = center_input - center_output
  246. # Calculate new composite affine
  247. new_affine_w_offset = self._create_comp_affine(affine, offset)
  248. affine_w_offset = np.dot(affine_w_offset, new_affine_w_offset)
  250. #----------------------------------------#
  251. # Scale along x, y axes, possibly z-axis #
  252. #----------------------------------------#
  253. if scale_xy:
  254. scale_factor = 1 + np.random.uniform(-scale_xy, scale_xy)
  255. affine = np.identity(3) * scale_factor
  257. # Eliminate scaling in the z-axis (may be preferable with
  258. # thick slices)
  259. if not scale_z:
  260. affine[2, 2] = 1
  261. else:
  262. affine[2, 2] = scale_z
  264. # Calculate center offset
  265. center_input = 0.5 * np.array(src.shape)
  266. center_output = center_input.dot(affine)
  267. offset = center_input - center_output
  269. # Calculate new composite affine
  270. new_affine_w_offset = self._create_comp_affine(affine, offset)
  271. affine_w_offset = np.dot(affine_w_offset, new_affine_w_offset)
  273. #------------------------------------------------#
  274. # Translate in the x and y axes, possibly z-axis #
  275. #------------------------------------------------#
  276. if translation_xy:
  277. x_offset = np.random.randint(-translation_xy,
  278. translation_xy + 1)
  279. y_offset = np.random.randint(-translation_xy,
  280. translation_xy + 1)
  281. else:
  282. x_offset, y_offset = 0, 0
  284. if translation_z:
  285. z_offset = np.random.randint(-translation_z, translation_z + 1)
  286. else:
  287. z_offset = 0
  289. offset = np.array([x_offset, y_offset, z_offset])
  291. # Add offset to existing affine offset
  292. affine_w_offset[:3, 3] = affine_w_offset[:3, 3] + offset
  294. #----------------------------------------#
  295. # Apply composite affine to src and mask #
  296. #----------------------------------------#
  297. output_img = ndimage.affine_transform(
  298. input=src,
  299. matrix=affine_w_offset[:3, :3].T,
  300. offset=affine_w_offset[:3, 3].ravel(),
  301. mode=filling_mode,
  302. cval=padding_value)
  303. output_mask = ndimage.affine_transform(
  304. input=mask,
  305. matrix=affine_w_offset[:3, :3].T,
  306. offset=affine_w_offset[:3, 3].ravel(),
  307. mode=filling_mode, cval=mask_PV )
  309. #---------------------#
  310. # Make thicker slices #
  311. #---------------------#
  312. # if len(thickness) > 1 or thickness[0] != 1:
  313. # th = np.random.choice(thickness)
  314. #
  315. # temp_output_img = np.zeros(output_img.shape[:2] + (output_img.shape[2] // th,))
  316. # temp_output_mask = np.zeros(output_img.shape[:2] + (output_img.shape[2] // th,))
  317. #
  318. # for i, j in enumerate(range(0, output_img.shape[2] // th * th, th)):
  319. # temp_output_img[:, :, i] = np.mean(output_img[:, :, j:j+th], axis=2)
  320. # temp_output_mask[:, :, i] = np.mean(output_mask[:, :, j:j+th], axis=2)
  321. #
  322. # output_img = temp_output_img
  323. # output_mask = temp_output_mask
  324. # else:
  325. # th = 1
  327. #---------------------------------------#
  328. # Flip image horizontally or vertically #
  329. #---------------------------------------#
  330. if flip_v:
  331. flip_v_bool = np.random.choice((True, False))
  332. if flip_v_bool:
  333. output_img = np.fliplr(output_img)
  334. output_mask = np.fliplr(output_mask)
  335. if flip_h:
  336. flip_h_bool = np.random.choice((True, False))
  337. if flip_h_bool:
  338. output_img = np.flipud(output_img)
  339. output_mask = np.flipud(output_mask)
  341. # Print augmentation parameters to screen
  342. # if verbose:
  343. # print(f'Aug {iteration+1}:')
  344. # if rotation_z:
  345. # print(' Rotation about z axis:',
  346. # f'{np.rad2deg(rot_z_angle):3} degrees')
  347. # if rotation_x:
  348. # print(' Rotation about x-axis:',
  349. # f'{np.rad2deg(rot_x_angle):3} degrees')
  350. # if rotation_y:
  351. # print(' Rotation about y-axis:',
  352. # f'{np.rad2deg(rot_y_angle):3} degrees')
  353. # if scale_xy:
  354. # print(f' Scaling in the x and y axes: {scale_factor:3}')
  355. # if scale_z:
  356. # print(f' Scaling in the z-axis: {scale_factor:3}')
  357. # if translation_xy:
  358. # print(f' X offset: {x_offset:3}')
  359. # if translation_z:
  360. # print(f' Y offset: {y_offset:3}')
  361. # if th != 1:
  362. # print(f' Slice thickness: {th}')
  363. # if flip_h:
  364. # print(f' Horizontal flip: {flip_h_bool}')
  365. # if flip_v:
  366. # print(f' Vertical flip: {flip_v_bool}')
  367. # print('')
  369. # Adjust for mask pixels that overlie bone
  370. if bone_edges:
  371. output_mask[np.logical_and(output_img > 100, output_mask < 0.85)] = 0
  373. # Make output_mask a binary mask again
  374. output_mask = (output_mask > 0.65).astype(float)
  376. # Concatenate augmentation with remainder of stack
  377. try:
  378. output_img_stack = np.concatenate(
  379. [output_img_stack, output_img], axis=2)
  380. output_mask_stack = np.concatenate(
  381. [output_mask_stack, output_mask], axis=2)
  382. except NameError:
  383. output_img_stack = output_img
  384. output_mask_stack = output_mask
  386. # if verbose:
  387. # print(f'Generated {n} augmentations(s) with a resulting' +
  388. # f' stack size of {output_img_stack.shape}.')
  390. return output_img_stack, output_mask_stack
  392. def _generate2d(self, src, mask, n, return_orig=True, verbose=0):
  393. """Generate `n` augmentations of a 2d `src` and `mask`.
  395. ### Args
  396. `src` (two dimensional ndarray): Image matrix.
  397. `mask` (two dimensional ndarray): Binary mask matrix.
  398. `n` (int): Number of augmentations.
  399. `return_orig` (bool): Determines whether original src and mask are
  400. returned. Defaults to True.
  401. `verbose` (int): Determines if augmentation parameters are printed
  402. to screen. Defaults to 0.
  404. ### Returns:
  405. Tuple of two matrices, first matrix contains stacked `src`
  406. augmentations while the second contains the stacked `mask`
  407. augmentations.
  408. """
  410. # Get parameters from self
  411. rotation_z = self.rotation_z
  412. rotation_x = self.rotation_x
  413. rotation_y = self.rotation_y
  414. translation_xy = self.translation_xy
  415. translation_z = self.translation_z
  416. scale_xy = self.scale_xy
  417. scale_z = self.scale_z
  418. flip_h = self.flip_h
  419. flip_v = self.flip_v
  420. filling_mode = self.filling_mode
  421. padding_value = self.padding_value
  422. mask_PV = self.mask_PV
  423. # Add original images and mask to output stack if warranted
  424. if return_orig:
  425. output_img_stack = src.reshape(src.shape + (1,)).copy()
  426. output_mask_stack = mask.reshape(mask.shape + (1,)).copy().astype(float)
  428. if verbose:
  429. print(self)
  431. for i in range(n):
  432. # Create base affine
  433. affine_w_offset = np.identity(3)
  435. #----------------------------------------------------#
  436. # Rotate around z-axis (rotation in the axial plane) #
  437. #----------------------------------------------------#
  438. if rotation_z:
  439. # Choose an angle at random and convert it to radians
  440. rot_z_angle = np.random.randint(-rotation_z, rotation_z + 1)
  441. rot_z_angle = np.deg2rad(rot_z_angle)
  443. affine = \
  444. np.array([[np.cos(rot_z_angle), -np.sin(rot_z_angle)],
  445. [np.sin(rot_z_angle), np.cos(rot_z_angle)]])
  447. # Calculate center offset
  448. center_input = 0.5 * np.array(src.shape)
  449. center_output = center_input.dot(affine)
  450. offset = center_input - center_output
  452. # Calculate new composite affine
  453. new_affine_w_offset = self._create_comp_affine(affine, offset)
  454. affine_w_offset = np.dot(affine_w_offset, new_affine_w_offset)
  456. #-----------------------#
  457. # Scale along x, y axes #
  458. #-----------------------#
  459. if scale_xy:
  460. scale_factor = 1 + np.random.uniform(-scale_xy, scale_xy)
  461. affine = np.identity(2) * scale_factor
  463. # Calculate center offset
  464. center_input = 0.5 * np.array(src.shape)
  465. center_output = center_input.dot(affine)
  466. offset = center_input - center_output
  468. # Calculate new composite affine
  469. new_affine_w_offset = self._create_comp_affine(affine, offset)
  470. affine_w_offset = np.dot(affine_w_offset, new_affine_w_offset)
  472. #------------------------------------------------#
  473. # Translate in the x and y axes, possibly z-axis #
  474. #------------------------------------------------#
  475. if translation_xy:
  476. x_offset = np.random.randint(-translation_xy,
  477. translation_xy + 1)
  478. y_offset = np.random.randint(-translation_xy,
  479. translation_xy + 1)
  480. else:
  481. x_offset, y_offset = 0, 0
  483. offset = np.array([x_offset, y_offset])
  485. # Add offset to existing affine offset
  486. affine_w_offset[:2, 2] = affine_w_offset[:2, 2] + offset
  488. #----------------------------------------#
  489. # Apply composite affine to src and mask #
  490. #----------------------------------------#
  491. output_img = ndimage.affine_transform(
  492. input=src,
  493. matrix=affine_w_offset[:2, :2].T,
  494. offset=affine_w_offset[:2, 2].ravel(),
  495. mode=filling_mode,
  496. cval=padding_value)
  497. output_mask = ndimage.affine_transform(
  498. input=mask,
  499. matrix=affine_w_offset[:2, :2].T,
  500. offset=affine_w_offset[:2, 2].ravel(),
  501. mode=filling_mode)
  503. #---------------------------------------#
  504. # Flip image horizontally or vertically #
  505. #---------------------------------------#
  506. if flip_h:
  507. flip_h_bool = np.random.choice((True, False))
  508. if flip_h_bool:
  509. output_img = np.fliplr(output_img)
  510. output_mask = np.fliplr(output_mask)
  511. if flip_v:
  512. flip_v_bool = np.random.choice((True, False))
  513. if flip_v_bool:
  514. output_img = np.flipud(output_img)
  515. output_mask = np.flipud(output_mask)
  517. # Print augmentation parameters to screen
  518. # if verbose:
  519. # print(f'Aug {i+1}:')
  520. # if rotation_z:
  521. # print(' Rotation about z axis:',
  522. # f'{np.rad2deg(rot_z_angle):3} degrees')
  523. # if rotation_x:
  524. # print(' Rotation about x-axis: 2d data, not performed')
  525. # if rotation_y:
  526. # print(' Rotation about y-axis: 2d data, not performed')
  527. # if scale_xy:
  528. # print(f' Scaling in the x and y axes: {scale_factor:3}')
  529. # if scale_z:
  530. # print(' Scaling in the z-axis: 2d data, not performed.')
  531. # if translation_xy:
  532. # print(f' X offset: {x_offset:3}')
  533. # if translation_z:
  534. # print(' Y offset: 2d data, not performed.')
  535. # if flip_h:
  536. # print(f' Horizontal flip: {flip_h_bool}')
  537. # if flip_v:
  538. # print(f' Vertical flip: {flip_v_bool}')
  539. # print('')
  541. # Make output_mask a binary mask again
  542. output_mask = (output_mask > 0.5).astype(int)
  544. # Convert output_img and output_mask to 3d matrices for concatenation
  545. output_img = output_img.reshape(output_img.shape + (1,))
  546. output_mask = output_mask.reshape(output_mask.shape + (1,))
  548. # Concatenate augmentation with remainder of stack
  549. try:
  550. output_img_stack = np.concatenate(
  551. [output_img_stack, output_img], axis=2)
  552. output_mask_stack = np.concatenate(
  553. [output_mask_stack, output_mask], axis=2)
  554. except NameError:
  555. output_img_stack = output_img
  556. output_mask_stack = (output_mask > 0.5).astype(int)
  558. # if verbose:
  559. # print(f'Generated {n} 2d augmentations(s) with a resulting' +
  560. # f' stack size of {output_img_stack.shape}.')
  562. return output_img_stack, output_mask_stack

auggen.py at commit d687482, no license · at the source

Overview

Authors: Zheng Sun1,2, Lifang Ye2, Yao Xu3, Yuehua Li2
  1. School of Health Science and Engineering, University of Shanghai for Science and Technology, Shanghai, China
  2. Institute of Diagnostic and Interventional Radiology, Shanghai Sixth People’s Hospital Affiliated to Shanghai Jiao Tong University School of Medicine, Shanghai, China
  3. Department of Radiology, Liyang Hospital of Chinese Medicine, Changzhou, Jiangsu, China
Institutions: University of Shanghai for Science and Technology (China); Shanghai Jiao Tong University (China); Shanghai Sixth People's Hospital (China)
Journal: Frontiers in neurology, volume 17, article 1895626
Dates: received 30 May 2026; accepted 20 July 2026; published online 3 August 2026
Type: Research article · Language: English
License: CC BY
Identifiers: DOI 10.3389/fneur.2026.1895626 · PMID 42609258 · PMCID PMC13477930 · OpenAlex W7172273600
Open access: gold, a free copy (OpenAlex)
Status: code verified
Categories: other (modality), human (organism), systems (subfield)
Methods: Statistics, Machine learning, Connectivity
Keywords: cross-center generalization, CT brain tissue extraction, deep learning, incremental fine-tuning, U-Net
MeSH: Brain*, Convolutional Neural Networks*, Tomography, X-Ray Computed*, Humans (* major topic)
Topic: Advanced Radiotherapy Techniques (Radiation, Physics and Astronomy), according to OpenAlex
Citations: not cited yet (Europe PMC); 24 references in the paper

Abstract

Background: To determine if a small number of annotated scans can enable clinically meaningful adaptation of a pre-trained convolutional neural network across different centers. Using non-contrast head CT brain extraction as the test case, we also sought to identify the factors contributing to performance drops when applying models to new sites.

Materials and methods: Non-contrast head CT scans were used in this study, from two separate institutions (Center A: 595 scans; Center B: 486 scans). A public pre-trained U-Net model for brain tissue extraction was first evaluated on Center A dataset as a baseline. The model was then fine-tuned incrementally using stratified subsets of annotated scans (5, 15, 25, and 75). The pre-trained model was subsequently transferred to Center B dataset and adapted through the same incremental fine-tuning strategy. The primary metric of interest was the three-dimensional Dice similarity coefficient (DSC), with a pre-defined minimal important difference (MID) set at 0.01. To identify predictors of baseline DSC and fine-tuning gain driving cross-center performance degradation, variance inflation factor (VIF) screening and logit Gaussian linear models (GLMs) were applied, with SHapley Additive exPlanations (SHAP) analysis providing qualitative feature importance rankings.

Results: At center A, the pre-trained model achieved a DSC of 0.9634 ± 0.0151. Fine-tuning with five annotated scans raised DSC to 0.9784 ± 0.0070 (p < 0.0001), exceeding the MID by 1.5-fold and reducing clinically unacceptable segmentations from 8.7 to 0.8%. A comparable gain was observed at Center B (p < 0.0001). Further increases to 15, 25, and 75 samples offered only marginal improvement below the MID threshold. Regression and SHAP analyses identified slice thickness as the dominant driver of both baseline performance degradation and fine-tuning improvement, accounting for 80–90% of predictive weight at both centers.

Conclusion: For NCCT brain extraction assessed across two centers, fine-tuning a publicly pre-trained CNN using five local annotated scans can generate clinically valuable cross-center segmentation gains; slice thickness heterogeneity appears to be the dominant source of inter-center domain shift.

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

Repository

Its files are read in the Code ↔ Paper reader above.

aqqush/CT_BET

License: none: the authors keep all their rights
State: the link answers, verified on 26 September 2026
Evidence: files inventoried
Commit: d6874828abf1e76fa398cbb86a36137dde120f50, 7 October 2020
Languages: Python (5)
Size: 14 files, 5 scripts
Software Heritage: not archived
Found in: the text, “Pre-trained model selection”
Holds: README
Not found: license file, CITATION.cff, environment file, tests, continuous integration, documentation
Tools: NumPy (4 files), NiBabel (3 files), Keras (2 files), SciPy (2 files), scikit-learn (1 file), SimpleITK (1 file)
Availability: 1 check, the latest on 26 September 2026: the link answers
  • 26 September 2026: the link answers
6 files
At the source: github.com/aqqush/CT_BET.git

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:

  • 1 repository of the authors' code, each at its verified commit, with its license and how the link was found in the paper;
  • 5 scripts, each with its path and the digest of its content;
  • no match between paragraphs and code yet;
  • 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 availability statement

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

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, 4 authors, 5 keywords, 4 MeSH terms, 23 references.

Cite

This paper

Sun, Z., Ye, L., Xu, Y., & Li, Y. (2026). Data-centric strategies enable cross-center generalization of convolutional neural networks for CT brain extraction. Frontiers in neurology, 17, 1895626. https://doi.org/10.3389/fneur.2026.1895626

BibTeX

@article{sun2026data,
author = {Sun, Zheng and Ye, Lifang and Xu, Yao and Li, Yuehua},
title = {{Data-centric strategies enable cross-center generalization of convolutional neural networks for CT brain extraction}},
journal = {Frontiers in neurology},
year = {2026},
month = aug,
volume = {17},
pages = {1895626},
publisher = {Frontiers Media SA},
issn = {1664-2295},
doi = {10.3389/fneur.2026.1895626},
url = {https://doi.org/10.3389/fneur.2026.1895626},
pmid = {42609258},
pmcid = {PMC13477930}
}

RIS

TY - JOUR
AU - Sun, Zheng
AU - Ye, Lifang
AU - Xu, Yao
AU - Li, Yuehua
TI - Data-centric strategies enable cross-center generalization of convolutional neural networks for CT brain extraction
T2 - Frontiers in neurology
J2 - Front Neurol
PY - 2026
DA - 2026/08/03
VL - 17
SP - 1895626
SN - 1664-2295
PB - Frontiers Media SA
DO - 10.3389/fneur.2026.1895626
UR - https://doi.org/10.3389/fneur.2026.1895626
LA - en
ER -

CSL-JSON

{
"id": "10.3389/fneur.2026.1895626",
"type": "article-journal",
"title": "Data-centric strategies enable cross-center generalization of convolutional neural networks for CT brain extraction",
"container-title": "Frontiers in neurology",
"author": [
{
"family": "Sun",
"given": "Zheng"
},
{
"family": "Ye",
"given": "Lifang"
},
{
"family": "Xu",
"given": "Yao"
},
{
"family": "Li",
"given": "Yuehua"
}
],
"container-title-short": "Front Neurol",
"volume": "17",
"page": "1895626",
"DOI": "10.3389/fneur.2026.1895626",
"PMID": "42609258",
"PMCID": "PMC13477930",
"ISSN": "1664-2295",
"publisher": "Frontiers Media SA",
"URL": "https://doi.org/10.3389/fneur.2026.1895626",
"language": "en",
"issued": {
"date-parts": [
[
2026,
8,
3
]
]
}
}

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:10.1080/07853890.2026.2685416 [code]
Pulmonary and cerebral damage in COVID-19 survivors: is there any association?
Journal: Annals of medicine
In common: SimpleITK, Keras, NiBabel, 2 other tools, other
[2] doi:10.3389/fmed.2026.1875760 [code]
Adaptive multi-stage domain unlearning for white-matter lesion segmentation.
Journal: Frontiers in medicine
In common: SimpleITK, NiBabel, scikit-learn, 2 other tools, 1 reference
[3] doi:10.3389/frai.2026.1771088 [code]
Few-shot deployment of pretrained MRI transformers in brain imaging tasks.
Journal: Frontiers in artificial intelligence
In common: SimpleITK, NiBabel, scikit-learn, 2 other tools, 1 reference
[4] doi:10.1007/s12021-026-09817-x [code]
Circle of Willis-Guided Localization for Simultaneous Detection and Classification of Large Vessel Occlusions in Brain CTA.
Journal: Neuroinformatics
In common: SimpleITK, NiBabel, scikit-learn, 2 other tools, other
[5] doi:10.1371/journal.pcbi.1014555 [code]
Body surface potential driven personalisation of electrophysiological digital twins in hypertrophic cardiomyopathy.
Journal: PLoS computational biology
In common: SimpleITK, NiBabel, scikit-learn, 2 other tools, other
[6] doi:10.64898/2026.05.06.26352540 [code]
Generating synthetic tau-PET scans in Alzheimer’s disease from MRI, blood biomarkers and demographics with deep learning
Journal: medRxiv (preprint)
In common: Keras, NiBabel, scikit-learn, 2 other tools, other
[7] doi:10.1038/s41467-026-76242-8 [code]
Whole-brain, all-optical interrogation of neuronal dynamics underlying gut and vascular interoception in zebrafish.
Journal: Nature communications
In common: SimpleITK, NiBabel, scikit-learn, 2 other tools, systems
[8] doi:10.3389/fnins.2026.1873417 [code]
Functional brain network alterations in a rat model of dental malocclusion.
Journal: Frontiers in neuroscience
In common: SimpleITK, NiBabel, scikit-learn, 2 other tools, systems
[9] doi:10.1364/boe.605322 [code]
Generalized plaque digitization framework for multi-dimensional mesoscopic images.
Journal: Biomedical optics express
In common: SimpleITK, Keras, scikit-learn, 2 other tools
[10] doi:10.3389/fnsys.2026.1822122 [code]
Convergence-divergence circuits for multimodal integration of innate and learned opponent valences.
Journal: Frontiers in systems neuroscience
In common: Keras, NiBabel, scikit-learn, 2 other tools, systems

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.