def get_instance_segmentation_metrics(a,
b,
is_binary=False,
thresholds=None,
**kwargs):
'''
Computes instance segmentation metric based on cellpose/stardist implementation.
https://cellpose.readthedocs.io/en/latest/api.html#cellpose.metrics.average_precision
'''
try:
from cellpose import metrics
except:
check_cellpose_installation()
from cellpose import metrics
# Find connected components in binary mask
if is_binary:
a = label_mask(a, **kwargs)
b = label_mask(b, **kwargs)
if thresholds is None:
#https://github.com/cocodataset/cocoapi/blob/master/PythonAPI/pycocotools/cocoeval.py
thresholds = np.linspace(.5,
0.95,
int(np.round((0.95 - .5) / .05)) + 1,
endpoint=True)
ap, tp, fp, fn = metrics.average_precision(a, b, threshold=thresholds)
return ap, tp, fp, fn
def suppfig_metrics(test_root, save_root, save_figure=False):
""" cyto performance measured with AJI and boundary precision """
ntest = len(glob(os.path.join(test_root, '*_img.tif')))
test_data = [io.imread(os.path.join(test_root, '%03d_img.tif'%i)) for i in range(ntest)]
test_labels = [io.imread(os.path.join(test_root, '%03d_masks.tif'%i)) for i in range(ntest)]
masks = []
aji = []
model_type = 'cyto'
masks.append(np.load(os.path.join(save_root, 'cellpose_%s_masks.npy'%model_type), allow_pickle=True))
masks.append(np.load(os.path.join(save_root, 'maskrcnn_%s_masks.npy'%model_type), allow_pickle=True))
masks.append(np.load(os.path.join(save_root, 'stardist_%s_masks.npy'%model_type), allow_pickle=True))
#masks.append(np.load(os.path.join(save_root, 'unet3_residual_on_style_on_concatenation_off_%s_masks.npy'%model_type), allow_pickle=True))
masks.append(np.load(os.path.join(save_root, 'unet3_residual_off_style_off_concatenation_on_%s_masks.npy'%model_type), allow_pickle=True))
for j in range(len(masks)):
aji.append(metrics.aggregated_jaccard_index(test_labels, masks[j]))
fsc.append(metrics.aggregated_jaccard_index(test_labels, masks[j]))
for m,model_type in enumerate(model_types):
masks[m].append(np.load(os.path.join(save_root, 'cellpose_%s_masks.npy'%model_type), allow_pickle=True))
masks[m].append(np.load(os.path.join(save_root, 'maskrcnn_%s_masks.npy'%model_type), allow_pickle=True))
masks[m].append(np.load(os.path.join(save_root, 'stardist_%s_masks.npy'%model_type), allow_pickle=True))
#masks[m].append(np.load(os.path.join(save_root, 'unet3_residual_on_style_on_concatenation_off_%s_masks.npy'%model_type), allow_pickle=True))
masks[m].append(np.load(os.path.join(save_root, 'unet3_residual_off_style_off_concatenation_on_%s_masks.npy'%model_type), allow_pickle=True))
masks[m].append(np.load(os.path.join(save_root, 'unet2_residual_off_style_off_concatenation_on_%s_masks.npy'%model_type), allow_pickle=True))
for j in range(len(masks[m])):
aps[m].append(metrics.average_precision(test_labels, masks[m][j],
threshold=thresholds)[0])
def test_unets(model_root, test_root, save_root, model_type='cyto'):
""" test trained unets """
device = mx.gpu()
ntest = len(glob(os.path.join(test_root, '*_img.tif')))
if model_type[:4] == 'cyto':
channels = [2, 1]
else:
channels = [0, 0]
concatenation = [1, 1, 0]
residual_on = [0, 0, 1]
style_on = [0, 0, 1]
nclasses = [3, 2, 3]
sstr = ['off', 'on']
aps = np.zeros((len(concatenation), ntest, len(thresholds)))
test_data = [
io.imread(os.path.join(test_root, '%03d_img.tif' % i))
for i in range(ntest)
]
test_labels = [
io.imread(os.path.join(test_root, '%03d_masks.tif' % i))
for i in range(ntest)
]
if model_type != 'cyto_sp':
dat = np.load(os.path.join(test_root, 'predicted_diams.npy'),
allow_pickle=True).item()
if model_type == 'cyto':
rescale = 30. / dat['predicted_diams']
else:
rescale = 17. / dat['predicted_diams']
else:
rescale = np.ones(len(test_data))
for k in range(1): #len(concatenation)):
pretrained_models = get_pretrained_models(model_root, 1, nclasses[k],
residual_on[k], style_on[k],
concatenation[k])
print(pretrained_models)
model = models.UnetModel(device=device,
pretrained_model=pretrained_models)
masks = model.eval(test_data,
channels=channels,
rescale=rescale,
net_avg=True)[0]
ap = metrics.average_precision(test_labels,
masks,
threshold=thresholds)[0]
print(ap[:, [0, 5, 8]].mean(axis=0))
aps[k] = ap
np.save(
os.path.join(
save_root,
'unet%d_residual_%s_style_%s_concatenation_%s_%s_masks.npy' %
(nclasses[k], sstr[residual_on[k]], sstr[style_on[k]],
sstr[concatenation[k]], model_type)), masks)
def cyto3d(save_root, save_figure=True):
thresholds = np.arange(0.25, 1.0, 0.05)
model_archs = ['ilastik',
'cellpose', 'unet3', 'unet2',
'cellpose_stitch', 'maskrcnn_stitch', 'stardist_stitch']
model_names = [['\u2014 ilastik'],
['\u2014 cellpose3D', '\u2014 unet3', '\u2014 unet2'],
['-- cellpose', '-- mask r-cnn', '-- stardist']]
titles = ['3D trained', '2D extended', '2D stitched']
colors = ['xkcd:periwinkle', 'm', 'y', 'r', 'm', 'g',
'c']
linestyles = ['-', '-', '-', '-', '--', '--', '--']
linewidths = [1,1,1,1,1,1,1]
masks_gt = np.load(os.path.join(save_root, 'ground_truth_3D_masks.npy'))
masks = []
aps = []
for model_arch in model_archs:
masks.append(np.load(os.path.join(save_root, '%s_3D_masks.npy'%model_arch)))
thresholds = np.arange(0.25,1.0,0.05)
aps.append(metrics.average_precision(masks_gt, masks[-1], threshold=thresholds)[0])
aps = np.array(aps)
print(aps[:,0])
ltrf = 10
rc('font', **{'size': 6})
fig = plt.figure(figsize=(2,2),facecolor='w',frameon=True, dpi=300)
ax = fig.add_subplot(111)
for m,model_arch in enumerate(model_archs):
ax.plot(thresholds, aps[m], colors[m],
linestyle=linestyles[m], linewidth=linewidths[m])
ax.spines['right'].set_visible(False)
ax.spines['top'].set_visible(False)
j=0
k=0
dy=0.08
for title,mnames in zip(titles,model_names):
ax.text(0.7, 0.95-dy*j, title, transform=ax.transAxes)
j+=1
for mname in mnames:
ax.text(0.7, 0.95-dy*j, mname, color=colors[k], transform=ax.transAxes)
k+=1
j+=1
ax.set_ylabel('average precision')
ax.set_xlabel('IoU matching threshold')
ax.set_xlim([0.25,1.0])
ax.set_ylim([0.,1.0])
if save_figure:
os.makedirs(os.path.join(save_root, 'figs'), exist_ok=True)
fig.savefig(os.path.join(save_root, 'figs/fig_perf3d.pdf'), bbox_inches='tight')
def test_cellpose_kfold_aug(data_root, save_root):
""" test trained cellpose networks on all cyto images """
device = mx.gpu()
ntest = 68
concatenation = [0]
residual_on = [1]
style_on = [1]
channels = [2, 1]
aps = np.zeros((9, 68, len(thresholds)))
for j in range(9):
train_root = os.path.join(data_root, 'train%d/' % j)
model_root = os.path.join(train_root, 'models/')
test_root = os.path.join(data_root, 'test%d/' % j)
test_data = [
io.imread(os.path.join(test_root, '%03d_img.tif' % i))
for i in range(ntest)
]
test_labels = [
io.imread(os.path.join(test_root, '%03d_masks.tif' % i))
for i in range(ntest)
]
k = 0
pretrained_models = get_pretrained_models(model_root, 0, 3,
residual_on[k], style_on[k],
concatenation[k])
print(pretrained_models)
cp_model = models.CellposeModel(device=device,
pretrained_model=pretrained_models)
dat = np.load(test_root + 'predicted_diams.npy',
allow_pickle=True).item()
rescale = 30. / dat['predicted_diams']
masks = cp_model.eval(test_data,
channels=channels,
rescale=rescale,
net_avg=True,
augment=True)[0]
ap = metrics.average_precision(test_labels,
masks,
threshold=thresholds)[0]
print(ap[:, [0, 5, 8]].mean(axis=0))
aps[j] = ap
return aps
def compare_masks(data_dir, image_names, runtype, model_type):
"""
Helper function to check if outputs given by a test are exactly the same
as the ground truth outputs.
"""
data_dir_2D = data_dir.joinpath('2D')
data_dir_3D = data_dir.joinpath('3D')
for image_name in image_names:
check = False
if '2D' in runtype and '2D' in image_name:
image_file = str(data_dir_2D.joinpath(image_name))
name = os.path.splitext(image_file)[0]
output_test = name + '_cp_masks.png'
output_true = name + '_%s_masks.png' % model_type
check = True
elif '3D' in runtype and '3D' in image_name:
image_file = str(data_dir_3D.joinpath(image_name))
name = os.path.splitext(image_file)[0]
output_test = name + '_cp_masks.tif'
output_true = name + '_%s_masks.tif' % model_type
check = True
if check:
if os.path.exists(output_test):
print('checking output %s' % output_test)
masks_test = io.imread(output_test)
masks_true = io.imread(output_true)
ap = metrics.average_precision(masks_true, masks_test)[0]
print('average precision of [%0.3f %0.3f %0.3f]' %
(ap[0], ap[1], ap[2]))
ap_precision = np.allclose(ap,
np.ones(3),
rtol=r_tol,
atol=a_tol)
matching_pix = np.logical_and(masks_test > 0,
masks_true > 0).mean()
all_pix = (masks_test > 0).mean()
pix_precision = np.allclose(all_pix,
matching_pix,
rtol=r_tol,
atol=a_tol)
assert all([ap_precision, pix_precision])
else:
print('ERROR: no output file of name %s found' % output_test)
assert False
def mask_stats(test_root, save_root, save_figure=False):
""" cyto performance broken down """
ntest = len(glob(os.path.join(test_root, '*_img.tif')))
test_data = [io.imread(os.path.join(test_root, '%03d_img.tif'%i)) for i in range(ntest)]
test_labels = [io.imread(os.path.join(test_root, '%03d_masks.tif'%i)) for i in range(ntest)]
masks = []
aps = []
model_type = 'cyto'
masks.append(np.load(os.path.join(save_root, 'cellpose_%s_masks.npy'%model_type), allow_pickle=True))
masks.append(np.load(os.path.join(save_root, 'maskrcnn_%s_masks.npy'%model_type), allow_pickle=True))
masks.append(np.load(os.path.join(save_root, 'stardist_%s_masks.npy'%model_type), allow_pickle=True))
masks.append(np.load(os.path.join(save_root, 'unet3_residual_off_style_off_concatenation_on_%s_masks.npy'%model_type), allow_pickle=True))
masks.append(np.load(os.path.join(save_root, 'unet2_residual_off_style_off_concatenation_on_%s_masks.npy'%model_type), allow_pickle=True))
for j in range(len(masks)):
aps.append(metrics.average_precision(test_labels, masks[j],
threshold=thresholds)[0])
# compute shape measure + bin it
convexities = np.zeros(0)
maskinds = np.zeros(0, 'int')
for i in range(len(test_labels)):
_,solidity, _ = utils.get_mask_stats(test_labels[i])
convexities = np.append(convexities, solidity)
maskinds = np.append(maskinds, i*np.ones(len(solidity), 'int'))
if i==10:
ispec = len(convexities)
bins = np.array([np.percentile(convexities, ip) for ip in np.linspace(0,100,4)])
digi = np.digitize(np.clip(convexities.copy(),bins[0]+.01, bins[-1]-0.01),
bins=bins) - 1
# compute IoU in shape bins
nbins = 3
iou_threshold = 0.5
ioub = np.zeros((len(masks), nbins))
ioub_exc = np.zeros((len(masks), nbins))
ioub_ste = np.zeros((len(masks), nbins))
for j in range(len(masks)):
iouall=np.zeros(0)
for i in range(len(test_labels)):
iou = metrics._intersection_over_union(test_labels[i], masks[j][i])[1:,1:]
n_min = min(iou.shape[0], iou.shape[1])
costs = -(iou >= 0.5).astype(float) - iou / (2*n_min)
true_ind, pred_ind = linear_sum_assignment(costs)
iout = np.zeros(test_labels[i].max())
iout[true_ind] = iou[true_ind,pred_ind]
iouall = np.append(iouall, iout)
for d in np.unique(digi):
iou_d = iouall[ispec:][digi[ispec:]==d]
ioub_exc[j,d] = (iou_d<=iou_threshold).mean()
iou_d = iou_d[iou_d > iou_threshold]
ioub[j,d] = iou_d.mean()
ioub_ste[j,d] = iou_d.std() / (digi==d).sum()
ltrf = 10
rc('font', **{'size': 6})
fig = plt.figure(figsize=(6.85,3.85),facecolor='w',frameon=True, dpi=300)
mdl = ['cellpose', 'mask r-cnn', 'stardist', 'unet3', 'unet2']
col ='mgcyr'
titles = ['Cells : fluorescent', 'Cells : nonfluorescent',
'Cell membranes', 'Microscopy : other', 'Non-microscopy']
s=0
inds = [np.arange(11,28,1,int), np.arange(28,33,1,int),
np.arange(33,42,1,int), np.arange(42,55,1,int),
np.arange(55,ntest,1,int)]
for t in range(len(inds)):
ax = fig.add_axes([0.1+.22*t,0.66,0.14,0.28])
for j in range(len(mdl)):
ax.plot(thresholds, aps[j][inds[t]].mean(axis=0), color=col[j])
#print(aps[0][j][:11].mean(axis=0)[0])
ax.spines['right'].set_visible(False)
ax.spines['top'].set_visible(False)
ax.set_ylim([0, 1])
ax.set_xlim([0.5, 1])
if t==0:
plt.ylabel('average precision')
ax.set_xlabel('IoU matching threshold')
ax.text(-0.2, 1.1, titles[t], fontsize = 7, ha='left', transform=ax.transAxes)
if t==0:
ax.text(-.4, 1.1, string.ascii_lowercase[s], fontsize = 10, transform=ax.transAxes)
s+=1
for j in range(len(mdl)):
ax.text(.5, .9 - .075*j, mdl[j], color=col[j], fontsize=6, transform=ax.transAxes)
# size figs
# size dist for all images
sz_dist = np.load(os.path.join(save_root, 'size_distribution.npy'))
# ap for all images
aps=np.load(os.path.join(save_root, 'ap_cellpose_all.npy'))
# example low-high masks
d=np.load(os.path.join(save_root, 'example_size_masks.npy'), allow_pickle=True).item()
msks = d['masks']
imgs = d['imgs']
inds = d['inds']
szs = sz_dist.flatten()
ap5 = aps[:,0,:,0].flatten()
r,p = stats.pearsonr(szs, ap5)
print(r,p)
xb = np.linspace(0,1,3)
bs = [300, 200]
for j in range(2):
ax = fig.add_axes([.07, .2-0.3*(j), 0.14,0.28])
#inds = np.nonzero(np.logical_and(sz_dist[0,:]>xb[j], sz_dist[0,:]<xb[j+1]))[0]
ic = inds[j]
maski = plot.mask_overlay(imgs[j], msks[j])
patch = plot.interesting_patch(msks[j], bsize=bs[j])
ax.imshow(maski[np.ix_(patch[0], patch[1])])
ax.axis('off')
if j==0:
ax.text(0.1, 1.05, 'low homogeneity', transform=ax.transAxes)
ax.text(-.15,1.25,string.ascii_lowercase[s],transform=ax.transAxes, fontsize=10)
ax.text(-.0,1.25,'within-image size variability',transform=ax.transAxes, fontsize=7)
s+=1
else:
ax.text(0.1, 1.05, 'high homogeneity', transform=ax.transAxes)
ax = fig.add_axes([0.285, 0.04, 0.18,0.4])
ax.scatter(szs, ap5, s=0.5)
slope, intercept = stats.linregress(szs, ap5)[:2]
lp = szs*slope + intercept
ax.plot(szs, lp, color='k', lw=1)
ax.scatter(szs[inds], ap5[inds], s=40, color='r', marker='+', lw=1)
ax.set_ylabel('AP @ IoU>0.5')
ax.set_xlabel('homogeneity of mask areas\n(25th / 75th percentile)')
ax.spines['right'].set_visible(False)
ax.spines['top'].set_visible(False)
ax.text(1.05, 0.63, 'r=%0.2f,\np=%0.3f'%(r, p), ha='right',
transform=ax.transAxes, fontsize=6)
#ax.text(-.25,1.1,string.ascii_lowercase[s],transform=ax.transAxes, fontsize=10)
#ax.set_title('Within-image size variability')
sstr = ['low', 'medium', 'high']
print(s)
for k in range(3):
ic = np.nonzero(np.logical_and(convexities<bins[k+1], convexities>bins[k]))[0]
np.random.seed(20)
inds = np.random.permutation(len(ic))
for l in range(24):
ax = fig.add_axes([0.51+0.035*(l%8), 0.44-(3.3*k+l//8)*0.06, 0.03, 0.05])
imask = ic[inds[l]]
im = maskinds[imask]
imask -= np.nonzero(maskinds==im)[0][0]
slices = find_objects(test_labels[im]==(imask+1))
ax.imshow(test_labels[im][slices[0]]==(imask+1), vmin=0, vmax=1, cmap='gray_r')
ax.axis('off')
if l==0:
ax.text(0.1, 1.3,'%s'%(sstr[k]), ha='left', transform=ax.transAxes)
if k==0:
ax.text(-0.5, 2, string.ascii_lowercase[s], transform=ax.transAxes, fontsize=10)
s+=1
ax.text(0.1, 2, 'convexity distributions', transform=ax.transAxes, fontsize=7)
dx=0.44
pbins = np.arange(0,3)
ax = fig.add_axes([0.42+dx, 0.04,.11,.4])
for j in range(len(masks)):
ax.plot(pbins+0.04*j, ioub_exc[j], color=col[j], lw=1)
ax.scatter(pbins+0.04*j, ioub_exc[j], color=col[j], s=5)
#if k==1 and s==0:
# ax.text(0.7,.95-0.06*(4-j), mdl[j], transform=ax.transAxes, color=col[j])
ax.set_xticks([0,1,2])
ax.set_xticklabels(['low', 'medium' , 'high'])
ax.set_ylabel('miss rate')
ax.set_xlabel('mask convexity')
ax.text(.7, 1.1, 'IoU threshold = 0.5', transform=ax.transAxes)
ax.set_ylim([0,1.])
ax.spines['right'].set_visible(False)
ax.spines['top'].set_visible(False)
#ax.set_title('generalized data', fontsize=7)
ax.text(-.5, 1.1, string.ascii_lowercase[s], transform=ax.transAxes, fontsize=10)
s+=1
ax = fig.add_axes([0.6+dx, 0.04, .11,.4])
for j in range(len(masks)):
ax.errorbar(pbins+0.04*j, ioub[j], ioub_ste[j],#np.abs(ioubi[k,j] - ioubi_ste[k,j].T),
color=col[j], lw=1)
ax.scatter(pbins+0.04*j, ioub[j], s=5, color=col[j])
ax.set_ylim([0.4, 1.0])
ax.spines['right'].set_visible(False)
ax.spines['top'].set_visible(False)
ax.set_ylabel('average IoU of true positives')
ax.set_xlabel('mask convexity')
ax.set_xticks([0,1,2])
ax.set_xticklabels(['low', 'medium' , 'high'])
if save_figure:
os.makedirs(os.path.join(save_root, 'figs'), exist_ok=True)
fig.savefig(os.path.join(save_root, 'figs/fig_stats.pdf'), bbox_inches='tight')
def suppfig_cellpose_params(test_root, save_root, save_figure=False):
ap_cellpose_all = np.load(os.path.join(save_root, 'ap_cellpose_all.npy'))
rc('font', **{'size': 6})
fig=plt.figure(figsize=(3,1.5), facecolor='w',frameon=True, dpi=300)
colors = [c for c in plt.get_cmap('Dark2').colors]
ap_compare = ap_cellpose_all[:,0,:,0].flatten()
netstr = ['style off', 'residual off', 'concatenation on', 'unet architecture', 'one net']
bmax = 0.15
dbin = 0.02
for i in range(4):#ap_cellpose_all.shape[1]-1):
ax = fig.add_axes([0.1+i*0.5, 0.1, 0.38, 0.75])
if i<5:
apt = ap_cellpose_all[:,i+1,:,0].flatten()
else:
apt = ap_cellpose_all[:,5:,:,0].mean(axis=1).flatten()
diffs = apt - ap_compare
ax.text(-.1, 1.1, netstr[i], transform=ax.transAxes , fontsize=7)
hb = ax.hist(np.clip(diffs, -1*bmax+dbin/2, bmax-dbin/2), bins=np.arange(-bmax, bmax+dbin, dbin),
color=colors[i])
max_counts = hb[0].max()*1.05
dm = np.mean(diffs)
p = stats.wilcoxon(diffs).pvalue
print(p)
nstars = np.array([p<0.05, p<0.01, p<0.001]).sum()
ax.scatter(dm, max_counts*1.025, marker='v', color=colors[i], s=10)
ax.text(dm, max_counts*1.1, '%0.3f'%dm+'*'*nstars, ha='center', fontsize=6)
ax.set_xlabel('difference in average precision')
if i==0:
ax.set_ylabel('# of test images')
ax.text(-.3, 1.1, string.ascii_lowercase[i], transform=ax.transAxes, fontsize=11)
ax.spines['right'].set_visible(False)
ax.spines['top'].set_visible(False)
ax.set_xlim([-bmax, bmax])
#ax.set_xticks(np.arange(-0.3,0.4,0.1))
ax.set_ylim([0, max_counts*1.25])
# performance without specialized images
ntest = len(glob(os.path.join(test_root, '*_img.tif')))
test_data = [io.imread(os.path.join(test_root, '%03d_img.tif'%i)) for i in range(ntest)]
test_labels = [io.imread(os.path.join(test_root, '%03d_masks.tif'%i)) for i in range(ntest)]
masks = []
aps = []
model_type = 'cyto'
masks.append(np.load(os.path.join(save_root, 'cellpose_%s_masks.npy'%model_type), allow_pickle=True))
masks.append(np.load(os.path.join(save_root, 'cellpose_%s_wo_sp_masks.npy'%model_type), allow_pickle=True))
for j in range(len(masks)):
aps.append(metrics.average_precision(test_labels, masks[j],
threshold=thresholds)[0])
i=4
ax = fig.add_axes([0.1+i*0.5, 0.1, 0.38, 0.75])
inds = [np.arange(11,ntest,1,int)]
sstr = ['cellpose', 'cellpose w/o\nspecialized']
ls = ['-']
col = ['m', [0.3, 0, 0.3]]
for t in range(len(inds)):
for j in range(len(masks)):
ax.plot(thresholds, aps[j][inds[t]].mean(axis=0), color=col[j], ls=ls[t])
if t==0:
ax.text(1., 0.7-j*0.18, sstr[j], color=col[j], ha='right')
#print(aps[0][j][:11].mean(axis=0)[0])
ax.spines['right'].set_visible(False)
ax.spines['top'].set_visible(False)
ax.set_ylim([0, 1])
ax.set_xlim([0.5, 1])
ax.set_ylabel('average precision')
ax.set_xlabel('IoU matching threshold')
ax.text(-.1, 1., 'remove specialized images\n from training set', transform=ax.transAxes , fontsize=7)
ax.text(-.3, 1.1, string.ascii_lowercase[i], transform=ax.transAxes, fontsize=11)
if save_figure:
os.makedirs(os.path.join(save_root, 'figs'), exist_ok=True)
fig.savefig(os.path.join(save_root, 'figs/suppfig_cellpose_params.pdf'),
bbox_inches='tight')
def cyto(test_root, save_root, save_figure=False):
""" cyto performance, main fig 2 """
ntest = len(glob(os.path.join(test_root, '*_img.tif')))
test_data = [io.imread(os.path.join(test_root, '%03d_img.tif'%i)) for i in range(ntest)]
test_labels = [io.imread(os.path.join(test_root, '%03d_masks.tif'%i)) for i in range(ntest)]
masks = [[], []]
aps = [[], []]
model_types = ['cyto_sp', 'cyto']
for m,model_type in enumerate(model_types):
masks[m].append(np.load(os.path.join(save_root, 'cellpose_%s_masks.npy'%model_type), allow_pickle=True))
masks[m].append(np.load(os.path.join(save_root, 'maskrcnn_%s_masks.npy'%model_type), allow_pickle=True))
masks[m].append(np.load(os.path.join(save_root, 'stardist_%s_masks.npy'%model_type), allow_pickle=True))
#masks[m].append(np.load(os.path.join(save_root, 'unet3_residual_on_style_on_concatenation_off_%s_masks.npy'%model_type), allow_pickle=True))
masks[m].append(np.load(os.path.join(save_root, 'unet3_residual_off_style_off_concatenation_on_%s_masks.npy'%model_type), allow_pickle=True))
masks[m].append(np.load(os.path.join(save_root, 'unet2_residual_off_style_off_concatenation_on_%s_masks.npy'%model_type), allow_pickle=True))
for j in range(len(masks[m])):
aps[m].append(metrics.average_precision(test_labels, masks[m][j],
threshold=thresholds)[0])
ltrf = 10
rc('font', **{'size': 6})
fig = plt.figure(figsize=(6.85,3.75/2 * 3),facecolor='w',frameon=True, dpi=300)
mdl = ['cellpose', 'mask r-cnn', 'stardist', 'unet3', 'unet2']
col ='mgcyr'
iimg = 1
for j in range(3):
ax = plt.subplot(3,4,2+j)
img = test_data[1][1]
img = np.stack((np.zeros_like(img), img, test_data[iimg][0]), axis=2)
plt.imshow(np.clip(img[:,75:-75,:], 0, 255))
outpix1 = utils.outlines_list(masks[0][j][iimg][:,75:-75])
outpix = utils.outlines_list(test_labels[iimg][:,75:-75])
for out in outpix:
plt.plot(out[:,0], out[:,1], color='y', lw=.5)
for out in outpix1:
plt.plot(out[:,0], out[:,1], '--', color='r', lw=.5)
plt.title(mdl[j], color=col[j], loc = 'left')
plt.text(.5, 1.05, '[email protected]=%.2f'%aps[0][j][iimg,0], transform=ax.transAxes, fontsize=7)
plt.arrow(305, 110, 7, 7, color='w', head_width = 5)
plt.arrow(255, 325, -10, 0, color='w', head_width = 5)
plt.arrow(155, 250, 10, 0, color='w', head_width = 5)
plt.arrow(315, 50, 0, -10, color='w', head_width = 5)
plt.arrow(100, 220, -7, -7, color='w', head_width = 5)
plt.axis('off')
if j==0:
plt.text(.0, 1.2, 'specialist model / specialized data', fontsize = 7, style='italic', transform=ax.transAxes)
plt.text(-.1, 1.2, 'b', fontsize = ltrf, transform=ax.transAxes)
iimg = 16
for j in range(3):
ax = plt.subplot(3,4,6+j)
img = test_data[iimg][1]
img = np.stack((img, img, img), axis=2)
plt.imshow(np.clip(img[:,:,:], 0, 255))
outpix1 = utils.outlines_list(masks[1][j][iimg][:,:])
outpix = utils.outlines_list(test_labels[iimg])
for out in outpix:
plt.plot(out[:,0], out[:,1], color='y', lw=.5)
for out in outpix1:
plt.plot(out[:,0], out[:,1], '--', color='r', lw=.5)
plt.title(mdl[j], color=col[j], loc = 'left', fontsize=7)
plt.text(.5, 1.1, '[email protected]=%.2f'%aps[1][j][iimg,0], transform=ax.transAxes, fontsize=6)
plt.axis('off')
if j==0:
plt.text(.0, 1.3, 'generalist model / generalized data', fontsize = 7, style='italic', transform=ax.transAxes)
plt.text(-.1, 1.3, 'c', fontsize = ltrf, transform=ax.transAxes)
titles = ['specialist model / \n specialized data', 'specialist model /\n generalized data',
'generalist model / \n specialized data', 'generalist model /\n generalized data']
inds = [np.arange(0,11,1,int), np.arange(11,ntest,1,int)]
for t in range(4):
ax = fig.add_axes([0.1+.22*t,0.1,0.17,0.25])
for j in range(len(mdl)):
ap = aps[t//2][j][inds[t%2]].mean(axis=0)
#print(titles[t], mdl[j], ap[0])
ax.plot(thresholds, ap, color=col[j])
#print(aps[0][j][:11].mean(axis=0)[0])
if t==2:
print(mdl[j], aps[t//2][j].mean(axis=0)[0])
ax.spines['right'].set_visible(False)
ax.spines['top'].set_visible(False)
ax.set_ylim([0, 1])
ax.set_xlim([0.5, 1])
if t==0:
plt.ylabel('average precision')
ax.set_xlabel('IoU matching threshold')
ax.text(0, 1.05, titles[t], fontsize = 7, ha='left', transform=ax.transAxes)
ax.text(-.25, 1.15, string.ascii_lowercase[t+3], fontsize = 10, transform=ax.transAxes)
if t==0:
for j in range(len(mdl)):
ax.text(.05, .4 - .075*j, mdl[j], color=col[j], fontsize=6, transform=ax.transAxes)
ax = fig.add_axes([.05,.37,.25,.65])
img = io.imread(os.path.join(save_root, 'figs/training_schematic_final.PNG'))
ax.imshow(img)
ax.axis('off')
ax.text(0, 1.09, 'a', fontsize = ltrf, transform=ax.transAxes)
if save_figure:
os.makedirs(os.path.join(save_root, 'figs'), exist_ok=True)
fig.savefig(os.path.join(save_root, 'figs/fig_perf2d_cyto.pdf'), bbox_inches='tight')
def nuclei(test_root, save_root, save_figure=False):
""" nuclei performance, suppfig """
ntest = len(glob(os.path.join(test_root, '*_img.tif')))
test_data = [io.imread(os.path.join(test_root, '%03d_img.tif'%i)) for i in range(ntest)]
test_labels = [io.imread(os.path.join(test_root, '%03d_masks.tif'%i)) for i in range(ntest)]
masks = []
aps = []
model_type = 'nuclei'
masks.append(np.load(os.path.join(save_root, 'cellpose_%s_masks.npy'%model_type), allow_pickle=True))
masks.append(np.load(os.path.join(save_root, 'maskrcnn_%s_masks.npy'%model_type), allow_pickle=True))
masks.append(np.load(os.path.join(save_root, 'stardist_%s_masks.npy'%model_type), allow_pickle=True))
#masks.append(np.load(os.path.join(save_root, 'unet3_residual_on_style_on_concatenation_off_%s_masks.npy'%model_type), allow_pickle=True))
masks.append(np.load(os.path.join(save_root, 'unet3_residual_off_style_off_concatenation_on_%s_masks.npy'%model_type), allow_pickle=True))
masks.append(np.load(os.path.join(save_root, 'unet2_residual_off_style_off_concatenation_on_%s_masks.npy'%model_type), allow_pickle=True))
for j in range(len(masks)):
aps.append(metrics.average_precision(test_labels, masks[j],
threshold=thresholds)[0])
ltrf = 10
rc('font', **{'size': 6})
fig = plt.figure(figsize=(6.85/2,3.85),facecolor='w',frameon=True, dpi=300)
mdl = ['cellpose', 'mask r-cnn', 'stardist', 'unet3', 'unet2']
col ='mgcyr'
iimg = 25
for j in range(3):
ax = plt.subplot(3,2,2*(1+j)-1)
img = test_data[iimg][1]
img = np.stack((img, img, img), axis=2)
plt.imshow(np.clip(img[:,:,:], 0, 255))
outpix1 = utils.outlines_list(masks[j][iimg])
outpix = utils.outlines_list(test_labels[iimg])
for out in outpix:
plt.plot(out[:,0], out[:,1], color='y', lw=.5)
for out in outpix1:
plt.plot(out[:,0], out[:,1], '--', color='r', lw=.5)
plt.title(mdl[j], color=col[j], loc = 'left')
plt.text(.5, 1.05, '[email protected]=%.2f'%aps[j][iimg,0], transform=ax.transAxes, fontsize=6)
plt.axis('off')
if j==0:
plt.text(-.1, 1.2, 'b', fontsize = ltrf, transform=ax.transAxes)
ax=fig.add_axes([.65, .3 ,.33,.4])
for j in range(len(mdl)):
ax.plot(thresholds, aps[j].mean(axis=0), color=col[j], lw=1.)
#print(aps[0][j][:11].mean(axis=0)[0])
ax.spines['right'].set_visible(False)
ax.spines['top'].set_visible(False)
ax.set_ylim([0, 1])
ax.set_xlim([0.5, 1])
ax.set_ylabel('average precision')
ax.set_xlabel('IoU matching threshold')
for j in range(len(mdl)):
ax.text(.05, .32 - .075*j, mdl[j], color=col[j], fontsize=6, transform=ax.transAxes)
ax.text(-.4, 1., 'c', fontsize = ltrf, transform=ax.transAxes)
if save_figure:
os.makedirs(os.path.join(save_root, 'figs'), exist_ok=True)
fig.savefig(os.path.join(save_root, 'figs/suppfig_perf2d_nuclei.pdf'), bbox_inches='tight')
return masks, aps