def sample_patches_retouch_mask(img,
mask=None,
pshape=(224, 224),
npos=10,
nneg=1,
pos=255,
neg=0,
patch_border=12):
"""
Generate patches from the interesting region of the OCT slice
:param img: oct image slice
:param mask: oct segmentation GT
:param pshape: patch shape
:param npos: Number of patches to sample from interesting region
:param nneg: Number of patches to sample from non interesting region
:param pos: Mask value indicating positives
:param neg: Mask value indicating negative
:param patch_border: boder to ignore when creating IRF,SRF,PED labels for patches (ignore border pixels for predicting labels)
:return:
"""
assert mask.dtype == np.uint8
roi_mask = np.logical_not(mask == 0)
it = ni.sample_patch_centers(roi_mask,
pshape=pshape,
npos=npos,
nneg=nneg,
pos=pos,
neg=neg)
for r, c, label in it:
img_patch = ni.extract_patch(img, pshape, r, c)
mask_patch = ni.extract_patch(mask, pshape, r, c)
label_IRF = np.int8(
np.any(mask_patch[patch_border:-patch_border,
patch_border:-patch_border] == IRF_CODE))
label_SRF = np.int8(
np.any(mask_patch[patch_border:-patch_border,
patch_border:-patch_border] == SRF_CODE))
label_PED = np.int8(
np.any(mask_patch[patch_border:-patch_border,
patch_border:-patch_border] == PED_CODE))
yield img_patch, mask_patch, label_IRF, label_SRF, label_PED
def test_extract_patch():
expected = np.array([[5, 6, 7], [9, 10, 11]], dtype='uint8')
image = np.reshape(np.arange(16, dtype='uint8'), (4, 4))
patch = ni.extract_patch(image, (2, 3), 2, 2)
nt.assert_allclose(expected, patch)
def sample_patches_entropy_mask(img,
mask=None,
roimask=None,
pshape=(224, 224),
npos=10,
nneg=1,
pos=255,
neg=0,
patch_border=12):
"""
Generate patches from the interesting region of the OCT slice
:param img: oct image slice
:param mask: oct segmentation GT
:param roimask: oct ROI
:param pshape: patch shape
:param npos: Number of patches to sample from interesting region
:param nneg: Number of patches to sample from non interesting region
:param pos: Mask value indicating positives
:param neg: Mask value indicating negative
:param patch_border: boder to ignore when creating IRF,SRF,PED labels for patches (ignore border pixels for predicting labels)
:return:
"""
# y_min, y_max = calculate_oct_y_range(img)
# roi_mask = np.zeros(mask.shape, dtype=np.int8)
# # print y_min, y_max
# roi_mask[y_min:y_max, :] = pos
# # TODO : Why 32?
# roi_mask[y_min:y_min + 32, :] = 0
# roi_mask[y_max - 32:y_max, :] = 0
# # plt.imshow(img)
# # print np.max(roi_mask), np.min(roi_mask)
roimask = roimask.astype(np.int8)
# it = ni.sample_patch_centers(roimask, pshape=pshape, npos=npos, nneg=nneg, pos=pos, neg=neg)
it = sample_oct_patch_centers(roimask,
pshape=pshape,
npos=npos,
pos=pos,
neg=neg)
# TODO : check code
it1 = ni.sample_patch_centers(roimask,
pshape=pshape,
npos=int(float(npos) * .2),
nneg=0,
pos=pos,
neg=neg)
for r, c, l in it1:
it.append([r, c, l])
# h,w = roimask.shape
# fig1 = plt.figure()
# ax1 = fig1.add_subplot(111, aspect='equal')
for r, c, label in it:
img_patch = ni.extract_patch(img, pshape, r, c)
mask_patch = ni.extract_patch(mask, pshape, r, c)
label_IRF = np.int8(
np.any(mask_patch[patch_border:-patch_border,
patch_border:-patch_border] == IRF_CODE))
label_SRF = np.int8(
np.any(mask_patch[patch_border:-patch_border,
patch_border:-patch_border] == SRF_CODE))
label_PED = np.int8(
np.any(mask_patch[patch_border:-patch_border,
patch_border:-patch_border] == PED_CODE))
yield img_patch, mask_patch, label_IRF, label_SRF, label_PED