def hd3dn(all_grp,all_card_gt,all_pred,all_gt,all_pnames,metric_axis,pprint=False,do_hd=0):
list(filter(lambda a: a != "0.0", all_pnames))
list(filter(lambda a: a != 0.0, all_pnames))
_,C = all_card_gt.shape
unique_patients = torch.unique(all_grp)
list(filter(lambda a: a != 0.0, unique_patients))
unique_patients = unique_patients[unique_patients != torch.ones_like(unique_patients)*666]
unique_patients = [u.item() for u in unique_patients]
batch_hd = torch.zeros((len(unique_patients), C))
for i, p in enumerate(unique_patients):
try:
bool_p = [int(re.split('_',re.split('slice',x.item())[1])[0])==p for x in all_pnames]
data = "whs"
except:
bool_p = [int(re.split('_',re.split('Subj_',x.item())[1])[0])==p for x in all_pnames]
data = "ivd"
slices_p = all_pnames[bool_p]
if do_hd >0:
all_gt_p = all_gt[bool_p,:]
all_pred_p = all_pred[bool_p,:]
sn_p = [int(re.split('_',x)[1]) for x in slices_p]
ord_p = np.argsort(sn_p)
label_gt = all_gt_p[ord_p,...]
label_pred = all_pred_p[ord_p,...]
hd_3d_var_vec= [None] * C
for j in range(0,C):
label_pred_c = numpy.copy(label_pred)
label_pred_c[label_pred_c!=j]=0
label_pred_c[label_pred_c==j]=1
label_gt_c = numpy.copy(label_gt)
label_gt_c[label_gt!=j]=0
label_gt_c[label_gt==j]=1
if len(np.unique(label_pred_c))>1 and len(np.unique(label_gt_c))>1:
if data=="whs":
hd_3d_var_vec[j] = hd95(label_pred_c, label_gt_c,[0.6,0.44,0.44]).item()
elif data=="ivd":
hd_3d_var_vec[j] = hd95(label_pred_c, label_gt_c,[1.25,1.25,2]).item()
else:
hd_3d_var_vec[j]=np.NaN
hd_3d_var=torch.from_numpy(np.asarray(hd_3d_var_vec))# np.nanmean(hd_3d_var_vec)
dice_3d = (2 * inter_card_p + 1e-8) / ((card_pred_p + card_gt_p)+ 1e-8)
if pprint and do_hd:
print(p,hd_3d_var)
if do_hd>0:
batch_hd[i,...] = hd_3d_var
indices = torch.tensor(metric_axis)
hd_3d = torch.index_select(batch_hd, 1, indices)
hd_3d_mean = hd_3d.mean(dim=0)
hd_3d_mean = torch.round(hd_3d_mean * 10**4) / (10**4)
hd_3d_sd = hd_3d.std(dim=0)
hd_3d_sd = torch.round(hd_3d_sd * 10**4) / (10**4)
print('metric_axis hd 3d',np.round(hd_3d_mean,2), 'mean ',np.round(hd_3d_mean.mean(),2),'std ',np.round(hd_3d_sd,2), 'std mean', np.round(hd_3d_sd.mean(),2))
[hd_3d, hd_3d_sd] = map_(lambda t: t.mean(), [hd_3d_mean, hd_3d_sd])
return hd_3d.item(), hd_3d_sd.item()
#Dice distance
dice = dc(mask_result, mask_reference)
#Jacard distance
jac = jc(mask_result, mask_reference)
#Haudorff distance
haus = hd(mask_result,
mask_reference,
voxelspacing=1.0156,
connectivity=1)
#Hausdorff distance 95
haus2 = hd95(mask_result,
mask_reference,
voxelspacing=1.0156,
connectivity=1)
fichier_dice.write(str(dice) + ';')
fichier_jac.write(str(jac) + ';')
fichier_haus.write(str(haus) + ';')
fichier_haus95.write(str(haus2) + ';')
fichier_dice.write("\n")
fichier_jac.write("\n")
fichier_haus.write("\n")
fichier_haus95.write("\n")
fichier_dice.close()
fichier_jac.close()
fichier_haus.close()
mdice.append(dice_value)
print(mpath, ': ', dice_value)
#other metrics
A = np.logical_and(img_pred[dice_idx], labels[dice_idx])
tp = float(A[A > 0].shape[0])
tn = float(A[A == 0].shape[0])
B = img_pred[dice_idx] - labels[dice_idx]
fp = float(B[B > 0].shape[0])
fn = float(B[B < 0].shape[0])
mspecificity.append(specificity(tp, tn, fp, fn))
msensitivity.append(sensitivity(tp, tn, fp, fn))
#mhausdorff.append(HausdorffDist(img_pred[dice_idx], labels[dice_idx]))
#mhausdorff.append(directed_hausdorff(img_pred[dice_idx], labels[dice_idx])[0])
mhausdorff.append(hd95(img_pred[dice_idx], labels[dice_idx]))
img_pred[dice_idx][img_pred[dice_idx] == 1] = 255
pred_dir = 'preds_' + ckpt_path[
7:] + '_reducedTest/' #'linknet_spp8/'
seg_path = pred_dir + os.path.basename(mpath[dice_idx])
if not os.path.exists(pred_dir):
os.mkdir(pred_dir)
for x in [30, 40, 50, 60, 70, 75, 80, 90, 100]:
pred_dir_sub = pred_dir + str(x)
if not os.path.exists(pred_dir_sub):
pass
os.mkdir(pred_dir_sub)
print('created', pred_dir_sub)
def dice3dn2(all_grp,all_inter_card,all_card_gt,all_card_pred,all_pred,all_gt,all_pnames,metric_axis,pprint=False,do_hd=0,best_epoch_val=0):
list(filter(lambda a: a != "0.0", all_pnames))
list(filter(lambda a: a != 0.0, all_pnames))
_,C = all_card_gt.shape
unique_patients = torch.unique(all_grp)
list(filter(lambda a: a != 0.0, unique_patients))
unique_patients = unique_patients[unique_patients != torch.ones_like(unique_patients)*666]
unique_patients = [u.item() for u in unique_patients]
batch_dice = torch.zeros((len(unique_patients), C))
batch_hd = torch.zeros((len(unique_patients), C))
batch_asd = torch.zeros((len(unique_patients), C))
if do_hd>0:
all_pred = all_pred.cpu().numpy()
all_gt = all_gt.cpu().numpy()
# do DICE
for i, p in enumerate(unique_patients):
inter_card_p = torch.einsum("bc->c", [torch.masked_select(all_inter_card, all_grp == p).reshape((-1, C))])
card_gt_p= torch.einsum("bc->c", [torch.masked_select(all_card_gt, all_grp == p).reshape((-1, C))])
card_pred_p= torch.einsum("bc->c", [torch.masked_select(all_card_pred, all_grp == p).reshape((-1, C))])
dice_3d = (2 * inter_card_p + 1e-8) / ((card_pred_p + card_gt_p)+ 1e-8)
batch_dice[i,...] = dice_3d
if pprint:
print(p,dice_3d.cpu())
indices = torch.tensor(metric_axis)
dice_3d = torch.index_select(batch_dice, 1, indices)
dice_3d_mean = dice_3d.mean(dim=0)
dice_3d_mean = torch.round(dice_3d_mean * 10**4) / (10**4)
dice_3d_sd = dice_3d.std(dim=0)
dice_3d_sd = torch.round(dice_3d_sd * 10**4) / (10**4)
# do HD
#if dice_3d_mean.mean()>best_epoch_val:
if dice_3d_mean.mean()>0:
for i, p in enumerate(unique_patients):
try:
#bool_p = [int(re.split('_',re.split('Case',x.item())[1])[0])==p for x in all_pnames]
bool_p = [int(re.split('_',re.split('slice',x.item())[1])[0])==p for x in all_pnames]
#bool_p = [int(re.split('_',re.split('Subj_',x.item())[1])[0])==p for x in all_pnames]
data = "saml"
except:
bool_p = [int(re.split('_',re.split('Case',x.item())[1])[0])==p for x in all_pnames]
#bool_p = [int(re.split('_',re.split('Subj_',x.item())[1])[0])==p for x in all_pnames]
data = "ivd"
slices_p = all_pnames[bool_p]
#if do_hd >0 or dice_3d_mean.mean()>best_epoch_val:
if do_hd >0 :
all_gt_p = all_gt[bool_p,:]
all_pred_p = all_pred[bool_p,:]
sn_p = [int(re.split('_',x)[1]) for x in slices_p]
ord_p = np.argsort(sn_p)
label_gt = all_gt_p[ord_p,...]
label_pred = all_pred_p[ord_p,...]
asd_3d_var_vec = [None] * C
hd_3d_var_vec= [None] * C
for j in range(0,C):
label_pred_c = numpy.copy(label_pred)
label_pred_c[label_pred_c!=j]=0
label_pred_c[label_pred_c==j]=1
label_gt_c = numpy.copy(label_gt)
label_gt_c[label_gt!=j]=0
label_gt_c[label_gt==j]=1
if len(np.unique(label_pred_c))>1 and len(np.unique(label_gt_c))>1:
if data=="saml":
asd_3d_var_vec[j] = assd(label_pred_c, label_gt_c).item() #ASD IN VOXEL TO COMPARE
hd_3d_var_vec[j] = asd_3d_var_vec[j]
elif data=="ivd":
hd_3d_var_vec[j] = hd95(label_pred_c, label_gt_c,[1.25,1.25,2]).item()
asd_3d_var_vec[j] = assd(label_pred_c, label_gt_c).item() #ASD IN VOXEL TO COMPARE TO PNP PAPER
else:
hd_3d_var_vec[j]=np.NaN
asd_3d_var_vec[j] = np.NaN
hd_3d_var=torch.from_numpy(np.asarray(hd_3d_var_vec))# np.nanmean(hd_3d_var_vec)
asd_3d_var=torch.from_numpy(np.asarray(asd_3d_var_vec))# np.nanmean(hd_3d_var_vec)
if pprint and do_hd:
print(p,asd_3d_var)
if do_hd>0:
batch_hd[i,...] = hd_3d_var
batch_asd[i,...] = asd_3d_var
hd_3d = torch.index_select(batch_hd, 1, indices)
asd_3d = torch.index_select(batch_asd, 1, indices)
hd_3d_mean = hd_3d.mean(dim=0)
asd_3d_mean = torch.from_numpy(np.nanmean(asd_3d, axis=0))
hd_3d_mean = torch.round(hd_3d_mean * 10**4) / (10**4)
asd_3d_mean = torch.round(asd_3d_mean * 10**4) / (10**4)
hd_3d_sd = hd_3d.std(dim=0)
asd_3d_sd = asd_3d.std(dim=0)
hd_3d_sd = torch.round(hd_3d_sd * 10**4) / (10**4)
asd_3d_sd = torch.round(asd_3d_sd * 10**4) / (10**4)
asd_3d_mean = asd_3d_mean.cpu().numpy()
#if dice_3d_mean.mean()<best_epoch_val:
#if dice_3d_mean.mean()<0:
# print('dice 3d',np.round(dice_3d_mean*100,2), 'mean ',np.round(100*dice_3d_mean.mean(),2),'std ',np.round(100*dice_3d_sd,2), 'std mean', np.round(100*dice_3d_sd.mean(),2))
#else:
print('dice 3d',np.round(dice_3d_mean*100,2), 'mean ',np.round(100*dice_3d_mean.mean(),2),'std ',np.round(100*dice_3d_sd,2), 'std mean', np.round(100*dice_3d_sd.mean(),2),'hd_3d',hd_3d_mean, 'hd std',hd_3d_sd,'asd',asd_3d_mean,'mean ',asd_3d_mean.mean())
[dice_3d, dice_3d_sd] = map_(lambda t: t.mean(), [dice_3d_mean, dice_3d_sd])
[hd_3d, hd_3d_sd] = map_(lambda t: t.mean(), [hd_3d_mean, hd_3d_sd])
[asd_3d, asd_3d_sd] = map_(lambda t: t.mean(), [asd_3d_mean, asd_3d_sd])
if pprint:
print('asd_3d_mean',asd_3d_mean.mean(), "asd_3d_sd", asd_3d_sd, "hd_3d_mean", hd_3d_mean, "hd_3d_sd", hd_3d_sd)
print("hd_3d_mean", hd_3d_mean, "hd_3d_sd", hd_3d_sd)
#print("in AA,LA,LV,Myo order:",asd_3d_sd.tolist()[3,1,2,0])
return dice_3d.item(), dice_3d_sd.item(), asd_3d.item(), asd_3d_sd.item()
def dice3dn(all_grp,all_inter_card,all_card_gt,all_card_pred,all_pred,all_gt,all_pnames,metric_axis,pprint=False,do_hd=0):
list(filter(lambda a: a != "0.0", all_pnames))
list(filter(lambda a: a != 0.0, all_pnames))
_,C = all_card_gt.shape
unique_patients = torch.unique(all_grp)
list(filter(lambda a: a != 0.0, unique_patients))
unique_patients = unique_patients[unique_patients != torch.ones_like(unique_patients)*666]
unique_patients = [u.item() for u in unique_patients]
batch_dice = torch.zeros((len(unique_patients), C))
batch_hd = torch.zeros((len(unique_patients), C))
batch_asd = torch.zeros((len(unique_patients), C))
if do_hd>0:
all_pred = all_pred.cpu().numpy()
all_gt = all_gt.cpu().numpy()
for i, p in enumerate(unique_patients):
try:
bool_p = [int(re.split('_',re.split('slice',x.item())[1])[0])==p for x in all_pnames]
data = "whs"
except:
bool_p = [int(re.split('_',re.split('Subj_',x.item())[1])[0])==p for x in all_pnames]
data = "ivd"
slices_p = all_pnames[bool_p]
inter_card_p = torch.einsum("bc->c", [torch.masked_select(all_inter_card, all_grp == p).reshape((-1, C))])
card_gt_p= torch.einsum("bc->c", [torch.masked_select(all_card_gt, all_grp == p).reshape((-1, C))])
card_pred_p= torch.einsum("bc->c", [torch.masked_select(all_card_pred, all_grp == p).reshape((-1, C))])
if do_hd >0:
all_gt_p = all_gt[bool_p,:]
all_pred_p = all_pred[bool_p,:]
sn_p = [int(re.split('_',x)[1]) for x in slices_p]
ord_p = np.argsort(sn_p)
label_gt = all_gt_p[ord_p,...]
label_pred = all_pred_p[ord_p,...]
hd_3d_var_vec= [None] * C
hd_3d_var_vec2= [None] * C
hd_3d_var_vec3= [None] * C
asd_3d_var_vec = [None] * C
for j in range(0,C):
label_pred_c = numpy.copy(label_pred)
label_pred_c[label_pred_c!=j]=0
label_pred_c[label_pred_c==j]=1
label_gt_c = numpy.copy(label_gt)
label_gt_c[label_gt!=j]=0
label_gt_c[label_gt==j]=1
if len(np.unique(label_pred_c))>1 and len(np.unique(label_gt_c))>1:
if data=="whs":
hd_3d_var_vec[j] = hd95(label_pred_c, label_gt_c,[0.6,0.44,0.44]).item()
asd_3d_var_vec[j] = assd(label_pred_c, label_gt_c).item() #ASD IN VOXEL TO COMPARE
elif data=="ivd":
hd_3d_var_vec[j] = hd95(label_pred_c, label_gt_c,[1.25,1.25,2]).item()
asd_3d_var_vec[j] = assd(label_pred_c, label_gt_c).item() #ASD IN VOXEL TO COMPARE TO PNP PAPER
else:
hd_3d_var_vec[j]=np.NaN
asd_3d_var_vec[j] = np.NaN
hd_3d_var=torch.from_numpy(np.asarray(hd_3d_var_vec))# np.nanmean(hd_3d_var_vec)
asd_3d_var=torch.from_numpy(np.asarray(asd_3d_var_vec))# np.nanmean(hd_3d_var_vec)
dice_3d = (2 * inter_card_p + 1e-8) / ((card_pred_p + card_gt_p)+ 1e-8)
if pprint and do_hd:
#dice_3d = torch.round(dice_3d * 10**4) / (10**4)
print(p,dice_3d.cpu(),hd_3d_var, asd_3d_var)
if pprint and not(do_hd):
print(p,dice_3d.cpu())
batch_dice[i,...] = dice_3d
if do_hd>0:
batch_hd[i,...] = hd_3d_var
batch_asd[i,...] = asd_3d_var
indices = torch.tensor(metric_axis)
dice_3d = torch.index_select(batch_dice, 1, indices)
hd_3d = torch.index_select(batch_hd, 1, indices)
asd_3d = torch.index_select(batch_asd, 1, indices)
dice_3d_mean = dice_3d.mean(dim=0)
hd_3d_mean = hd_3d.mean(dim=0)
asd_3d_mean = asd_3d.mean(dim=0)
dice_3d_mean = torch.round(dice_3d_mean * 10**4) / (10**4)
hd_3d_mean = torch.round(hd_3d_mean * 10**4) / (10**4)
asd_3d_mean = torch.round(asd_3d_mean * 10**4) / (10**4)
dice_3d_sd = dice_3d.std(dim=0)
hd_3d_sd = hd_3d.std(dim=0)
asd_3d_sd = asd_3d.std(dim=0)
dice_3d_sd = torch.round(dice_3d_sd * 10**4) / (10**4)
hd_3d_sd = torch.round(hd_3d_sd * 10**4) / (10**4)
asd_3d_sd = torch.round(asd_3d_sd * 10**4) / (10**4)
print('metric_axis dice 3d',np.round(dice_3d_mean*100,2), 'mean ',np.round(100*dice_3d_mean.mean(),2),'std ',np.round(100*dice_3d_sd,2), 'std mean', np.round(100*dice_3d_sd.mean(),2))
if pprint:
print('metric_axis asd 3d',np.round(asd_3d_mean,2), 'mean ',np.round(asd_3d_mean.mean(),2),'std ',np.round(asd_3d_sd,2), 'std mean', np.round(asd_3d_sd.mean(),2))
print('metric_axis hd 3d',np.round(hd_3d_mean,2), 'mean ',np.round(hd_3d_mean.mean(),2),'std ',np.round(hd_3d_sd,2), 'std mean', np.round(hd_3d_sd.mean(),2))
print("DICE",np.round(dice_3d_mean[0].item()*100,1), " $\pm$ ", np.round(dice_3d_sd[0].item()*100,1),
"&",np.round(dice_3d_mean[1].item()*100,1), " $\pm$ ", np.round(dice_3d_sd[1].item()*100,1),
"&",np.round(dice_3d_mean[2].item()*100,1), " $\pm$ ", np.round(dice_3d_sd[2].item()*100,1),
"&",np.round(dice_3d_mean[3].item()*100,1), " $\pm$ ", np.round(dice_3d_sd[3].item()*100,1),
"&",np.round(dice_3d_mean.mean().item()*100,1), " $\pm$ ", np.round(dice_3d_sd.mean().item()*100,1))
print("HD", np.round(hd_3d_mean[0].item(),1), " $\pm$ ", np.round(hd_3d_sd[0].item(),1),
"&",np.round(hd_3d_mean[1].item(),1), " $\pm$ ", np.round(hd_3d_sd[1].item(),1),
"&",np.round(hd_3d_mean[2].item(),1), " $\pm$ ", np.round(hd_3d_sd[2].item(),1),
"&",np.round(hd_3d_mean[3].item(),1), " $\pm$ ", np.round(hd_3d_sd[3].item(),1),
"&",np.round(hd_3d_mean.mean().item(),1), " $\pm$ ", np.round(hd_3d_sd.mean().item(),1))
print("ASD", np.round(asd_3d_mean[0].item(),1), " $\pm$ ", np.round(asd_3d_sd[0].item(),1),
"&",np.round(asd_3d_mean[1].item(),1), " $\pm$ ", np.round(asd_3d_sd[1].item(),1),
"&",np.round(asd_3d_mean[2].item(),1), " $\pm$ ", np.round(asd_3d_sd[2].item(),1),
"&",np.round(asd_3d_mean[3].item(),1), " $\pm$ ", np.round(asd_3d_sd[3].item(),1),
"&",np.round(asd_3d_mean.mean().item(),1), " $\pm$ ", np.round(asd_3d_sd.mean().item(),1))
[dice_3d, dice_3d_sd] = map_(lambda t: t.mean(), [dice_3d_mean, dice_3d_sd])
[hd_3d, hd_3d_sd] = map_(lambda t: t.mean(), [hd_3d_mean, hd_3d_sd])
[asd_3d, asd_3d_sd] = map_(lambda t: t.mean(), [asd_3d_mean, asd_3d_sd])
if pprint:
print('asd_3d_mean',asd_3d_mean, "asd_3d_sd", asd_3d_sd, "hd_3d_mean", hd_3d_mean, "hd_3d_sd", hd_3d_sd)
return dice_3d.item(), dice_3d_sd.item(), hd_3d.item(), hd_3d_sd.item()
def dice3d(all_grp,all_inter_card,all_card_gt,all_card_pred,all_pred,all_gt,all_pnames,metric_axis,pprint=False,do_hd=0,do_asd=0,best_epoch_val=0):
_,C = all_card_gt.shape
unique_patients = torch.unique(all_grp)
list(filter(lambda a: a != 0.0, unique_patients))
unique_patients = [u.item() for u in unique_patients]
batch_dice = torch.zeros((len(unique_patients), C))
batch_hd = torch.zeros((len(unique_patients), C))
batch_asd = torch.zeros((len(unique_patients), C))
if do_hd>0 or do_asd>0:
all_pred = all_pred.cpu().numpy()
all_gt = all_gt.cpu().numpy()
# do DICE
for i, p in enumerate(unique_patients):
inter_card_p = torch.einsum("bc->c", [torch.masked_select(all_inter_card, all_grp == p).reshape((-1, C))])
card_gt_p= torch.einsum("bc->c", [torch.masked_select(all_card_gt, all_grp == p).reshape((-1, C))])
card_pred_p= torch.einsum("bc->c", [torch.masked_select(all_card_pred, all_grp == p).reshape((-1, C))])
dice_3d = (2 * inter_card_p + 1e-8) / ((card_pred_p + card_gt_p)+ 1e-8)
batch_dice[i,...] = dice_3d
if pprint:
print(p,dice_3d.cpu())
indices = torch.tensor(metric_axis)
dice_3d = torch.index_select(batch_dice, 1, indices)
dice_3d_mean = dice_3d.mean(dim=0)
dice_3d_mean = torch.round(dice_3d_mean * 10**4) / (10**4)
dice_3d_sd = dice_3d.std(dim=0)
dice_3d_sd = torch.round(dice_3d_sd * 10**4) / (10**4)
# do HD and / or ASD
#if dice_3d_mean.mean()>best_epoch_val:
if dice_3d_mean.mean()>0:
for i, p in enumerate(unique_patients):
root_name = [re.split('(\d+)', x.item())[0] for x in all_pnames][0]
bool_p = [int(re.split('_',re.split(root_name,x.item())[1])[0])==p for x in all_pnames]
slices_p = all_pnames[bool_p]
#if do_hd >0 or dice_3d_mean.mean()>best_epoch_val:
if do_hd> 0 or do_asd >0 :
all_gt_p = all_gt[bool_p,:]
all_pred_p = all_pred[bool_p,:]
sn_p = [int(re.split('_',x)[1]) for x in slices_p]
ord_p = np.argsort(sn_p)
label_gt = all_gt_p[ord_p,...]
label_pred = all_pred_p[ord_p,...]
asd_3d_var_vec = [None] * C
hd_3d_var_vec= [None] * C
for j in range(0,C):
label_pred_c = numpy.copy(label_pred)
label_pred_c[label_pred_c!=j]=0
label_pred_c[label_pred_c==j]=1
label_gt_c = numpy.copy(label_gt)
label_gt_c[label_gt!=j]=0
label_gt_c[label_gt==j]=1
if len(np.unique(label_pred_c))>1: # len(np.unique(label_gt_c))>1 should always be true...
if do_hd>0:
if root_name=="Case" or root_name=="slice":
hd_3d_var_vec[j] = hd95(label_pred_c, label_gt_c,[0.6,0.44,0.44]).item()
elif root_name=="Subj_":
hd_3d_var_vec[j] = hd95(label_pred_c, label_gt_c,[1.25,1.25,2]).item()
if do_asd > 0:
asd_3d_var_vec[j] = assd(label_pred_c, label_gt_c).item() #ASD IN VOXEL TO COMPARE TO PNP PAPER
else:
#print(len(np.unique(label_pred_c)),len(np.unique(label_gt_c)))
hd_3d_var_vec[j]=np.NaN
asd_3d_var_vec[j] = np.NaN
if do_hd>0:
hd_3d_var = torch.from_numpy(np.asarray(hd_3d_var_vec)) # np.nanmean(hd_3d_var_vec)
batch_hd[i,...] = hd_3d_var
if do_asd>0:
asd_3d_var = torch.from_numpy(np.asarray(asd_3d_var_vec)) # np.nanmean(hd_3d_var_vec)
batch_asd[i,...] = asd_3d_var
[hd_3d, hd_3d_sd] = get_mean_sd(batch_hd,indices)
[asd_3d, asd_3d_sd] = get_mean_sd(batch_asd,indices)
[dice_3d, dice_3d_sd] = map_(lambda t: t.mean(), [dice_3d_mean.cpu().numpy(), dice_3d_sd.cpu().numpy()])
if pprint:
print('asd_3d_mean',asd_3d, "asd_3d_sd", asd_3d_sd, "hd_3d_mean", hd_3d, "hd_3d_sd", hd_3d_sd)
[return_asd,return_asd_sd] = [asd_3d.item(),asd_3d_sd.item()] if do_asd >0 else [0,0]
[return_hd,return_hd_sd] = [hd_3d.item(),hd_3d_sd.item()] if do_hd >0 else [0,0]
return dice_3d.item(), dice_3d_sd.item(), return_asd, return_asd_sd,return_hd,return_hd_sd
def dice3dn(all_grp,
all_inter_card,
all_card_gt,
all_card_pred,
all_pred,
all_gt,
all_pnames,
metric_axis,
pprint=False,
do_hd=0):
#print(all_card_gt.shape)
list(filter(lambda a: a != "0.0", all_pnames))
list(filter(lambda a: a != 0.0, all_pnames))
#print(all_pnames,"all_pnames")
_, C = all_card_gt.shape
unique_patients = torch.unique(all_grp)
list(filter(lambda a: a != 0.0, unique_patients))
#unique_patients = unique_patients[unique_patients!="0.0"]
#unique_patients = unique_patients[unique_patients!=0.0]
#print(sum(unique_patients == 0))
#print(all_pred.sum())
#print(all_gt.sum())
#print(np.unique(all_pred),np.unique(all_gt),"np.unique(all_pred),np.unique(all_gt)")
unique_patients = unique_patients[
unique_patients != torch.ones_like(unique_patients) * 666]
unique_patients = [u.item() for u in unique_patients]
#unique_patients = unique_patients[unique_patients != 666]
#print(unique_patients)
batch_dice = torch.zeros((len(unique_patients), C))
batch_hd = torch.zeros((len(unique_patients), C))
batch_asd = torch.zeros((len(unique_patients), C))
#hd_3d = torch.zeros((len(unique_patients)))
#hd95_3d = torch.zeros((len(unique_patients)))
#hd95_3d_bis = torch.zeros((len(unique_patients)))
#hd_3d_var = torch.zeros((len(unique_patients)))
#asd_3d = torch.zeros((len(unique_patients)))
#ravd_3d = torch.zeros((len(unique_patients)))
#print(len(all_pred))
if do_hd > 0:
all_pred = all_pred.cpu().numpy()
all_gt = all_gt.cpu().numpy()
for i, p in enumerate(unique_patients):
#print(p)
try:
bool_p = [
int(re.split('_',
re.split('slice', x.item())[1])[0]) == p
for x in all_pnames
]
data = "whs"
except:
bool_p = [
int(re.split('_',
re.split('Subj_', x.item())[1])[0]) == p
for x in all_pnames
]
data = "ivd"
#print(bool_p,"bool_p")
#print(np.sum(bool_p))
#bool_p = bool_p.detach()
#print(all_pnames[1:10])
#print(len(np.unique(all_pnames)))
#print(bool_p[1:10])
#print(all_grp[:,1])
slices_p = all_pnames[bool_p]
#print(all_pred.shape,all_grp.shape, "pred versus all_grp")
#print((all_grp[:,1] == p).shape,'all_grp[:,1] == p shape')
#print(all_grp,'all_grp')
#print(sum(all_grp[:,1] == p),'all_grp[:,1] == p sum')
#all_gt_p = all_gt[all_grp[:,1] == p,:,:]
#all_gt_p = all_gt[bool_p,:]
#all_pred_p = all_pred[all_grp[:,1] == p,:,:].cpu()
#all_pred_p = all_pred[bool_p,:]
#print(all_gt,all_pred)
#print(all_gt_p.shape,len(slices_p)) # should be the same
#sn_p = [int(re.split('_',x)[2].split('.')[0]) for x in slices_p]
#ord_p = np.argsort(sn_p)
#label_gt = all_gt_p[ord_p,...]
#print(all_gt_p, all_pred_p)
#label_pred = all_pred_p[ord_p,...]
#print(label_gt.sum())
#print(all_gt_p, all_pred_p)
inter_card_p = torch.einsum("bc->c", [
torch.masked_select(all_inter_card, all_grp == p).reshape((-1, C))
])
card_gt_p = torch.einsum(
"bc->c",
[torch.masked_select(all_card_gt, all_grp == p).reshape((-1, C))])
card_pred_p = torch.einsum("bc->c", [
torch.masked_select(all_card_pred, all_grp == p).reshape((-1, C))
])
#all_pred = all_pred.cpu().numpy()
#all_gt = all_gt.cpu().numpy()
#hd_3d[i]= hd(label_pred, label_gt,[1.25,1.25,2] ).item()
if do_hd > 0:
#print(slices_p)
#all_pred = all_pred.cpu().numpy()
#print(all_pred.shape)
#all_gt = all_gt.cpu().numpy()
all_gt_p = all_gt[bool_p, :]
all_pred_p = all_pred[bool_p, :]
#sn_p = [int(re.split('_',x)[2].split('.')[0]) for x in slices_p]
#print(slices_p)
#sn_p = [int(re.split('_',re.split('slice',x)[1])[0]) for x in slices_p]
sn_p = [int(re.split('_', x)[1]) for x in slices_p]
#print(sn_p)
ord_p = np.argsort(sn_p)
#print(ord_p)
label_gt = all_gt_p[ord_p, ...]
label_pred = all_pred_p[ord_p, ...]
#print(np.unique(label_pred))
hd_3d_var_vec = [None] * C
hd_3d_var_vec2 = [None] * C
hd_3d_var_vec3 = [None] * C
asd_3d_var_vec = [None] * C
#hd_3d_var_vec[:] = np.NaN
#asd_3d_vec[:] = np.NaN
#print(np.unique(label_pred), np.unique(label_gt))
for j in range(0, C):
#print(j)
#print(np.unique(label_pred),"j",j, np.unique(label_pred))
label_pred_c = numpy.copy(label_pred)
label_pred_c[label_pred_c != j] = 0
label_pred_c[label_pred_c == j] = 1
label_gt_c = numpy.copy(label_gt)
label_gt_c[label_gt != j] = 0
label_gt_c[label_gt == j] = 1
#print(np.unique(label_pred_c),np.unique(label_gt_c),"np.unique(label_pred_c),np.unique(label_gt_c)")
#hd_3d_var_vec[j-1] = hd_var(label_pred_c, label_gt_c,[1.25,1.25,2],1,do_hd).item()
if len(np.unique(label_pred_c)) > 1 and len(
np.unique(label_gt_c)) > 1:
if data == "whs":
#print("data:whs")
hd_3d_var_vec[j] = hd95(label_pred_c, label_gt_c,
[0.6, 0.44, 0.44]).item()
#hd_3d_var_vec2[j] = hd(label_pred_c, label_gt_c,[0.6,0.44,0.44]).item()
#hd_3d_var_vec3[j] = hd(label_pred_c, label_gt_c,[1,1,1]).item()
asd_3d_var_vec[j] = assd(
label_pred_c,
label_gt_c).item() #ASD IN VOXEL TO COMPARE
#print("hd100:",round(hd(label_pred_c, label_gt_c,[0.6,0.44,0.44]).item()), ", hd100 in vox",round(hd(label_pred_c, label_gt_c,[1,1,1]).item()), ", hd95:",round(hd_3d_var_vec[j]))
elif data == "ivd":
#print("data:ivd")
hd_3d_var_vec[j] = hd95(label_pred_c, label_gt_c,
[1.25, 1.25, 2]).item()
asd_3d_var_vec[j] = assd(
label_pred_c, label_gt_c).item(
) #ASD IN VOXEL TO COMPARE TO PNP PAPER
#print(hd_var(label_pred_c, label_gt_c,[1.25,1.25,2],1,do_hd).item())
#asd_3d_var_vec[j] = assd(label_pred_c, label_gt_c).item()
else:
hd_3d_var_vec[j] = np.NaN
asd_3d_var_vec[j] = np.NaN
#print("hd_3d_var_vec", hd_3d_var_vec)
hd_3d_var = torch.from_numpy(
np.asarray(hd_3d_var_vec)) # np.nanmean(hd_3d_var_vec)
#hd_3d_var2=torch.from_numpy(np.asarray(hd_3d_var_vec2))# np.nanmean(hd_3d_var_vec)
#hd_3d_var3=torch.from_numpy(np.asarray(hd_3d_var_vec3))# np.nanmean(hd_3d_var_vec)
asd_3d_var = torch.from_numpy(
np.asarray(asd_3d_var_vec)) # np.nanmean(hd_3d_var_vec)
#asd_3d[i]=np.nanmean(asd_3d_vec)
#hd95_3d_bis[i]= hd95(label_pred, label_gt,[1.25,2,1.25]).item()
#ravd_3d[i]= abs(ravd(label_pred, label_gt))
#print(hd_3d[i])
#if p == 0:
# print("inter_card_p:",inter_card_p.detach())
# print("card_gt_p:", card_gt_p.detach())
# print("card_pred_p:",card_pred_p.detach())
#print(card_gt_p.shape)
dice_3d = (2 * inter_card_p + 1e-8) / (
(card_pred_p + card_gt_p) + 1e-8)
if pprint and do_hd:
#dice_3d = torch.round(dice_3d * 10**4) / (10**4)
print(p, dice_3d.cpu(), hd_3d_var, asd_3d_var)
if pprint and not (do_hd):
print(p, dice_3d.cpu())
batch_dice[i, ...] = dice_3d
if do_hd > 0:
batch_hd[i, ...] = hd_3d_var
batch_asd[i, ...] = asd_3d_var
indices = torch.tensor(metric_axis)
dice_3d = torch.index_select(batch_dice, 1, indices)
hd_3d = torch.index_select(batch_hd, 1, indices)
asd_3d = torch.index_select(batch_asd, 1, indices)
dice_3d_mean = dice_3d.mean(dim=0)
hd_3d_mean = hd_3d.mean(dim=0)
asd_3d_mean = asd_3d.mean(dim=0)
dice_3d_mean = torch.round(dice_3d_mean * 10**4) / (10**4)
hd_3d_mean = torch.round(hd_3d_mean * 10**4) / (10**4)
asd_3d_mean = torch.round(asd_3d_mean * 10**4) / (10**4)
dice_3d_sd = dice_3d.std(dim=0)
hd_3d_sd = hd_3d.std(dim=0)
asd_3d_sd = asd_3d.std(dim=0)
dice_3d_sd = torch.round(dice_3d_sd * 10**4) / (10**4)
hd_3d_sd = torch.round(hd_3d_sd * 10**4) / (10**4)
asd_3d_sd = torch.round(asd_3d_sd * 10**4) / (10**4)
#hd95_3d_sd = hd95_3d.std()
#hd95_3d_mean = hd95_3d.mean()
#hdvar_3d_mean = hd_3d_var.mean()
#hdvar_3d_sd = hd_3d_var.std()
#ravd_3d_sd = ravd_3d.std()
#ravd_3d_mean = ravd_3d.mean()
print('metric_axis dice 3d', np.round(dice_3d_mean * 100, 2), 'mean ',
np.round(100 * dice_3d_mean.mean(), 2), 'std ',
np.round(100 * dice_3d_sd, 2), 'std mean',
np.round(100 * dice_3d_sd.mean(), 2))
if pprint:
print('metric_axis asd 3d', np.round(asd_3d_mean, 2), 'mean ',
np.round(asd_3d_mean.mean(), 2), 'std ', np.round(asd_3d_sd, 2),
'std mean', np.round(asd_3d_sd.mean(), 2))
print('metric_axis hd 3d', np.round(hd_3d_mean, 2), 'mean ',
np.round(hd_3d_mean.mean(), 2), 'std ', np.round(hd_3d_sd, 2),
'std mean', np.round(hd_3d_sd.mean(), 2))
print("DICE", np.round(dice_3d_mean[0].item() * 100, 1), " $\pm$ ",
np.round(dice_3d_sd[0].item() * 100, 1), "&",
np.round(dice_3d_mean[1].item() * 100, 1), " $\pm$ ",
np.round(dice_3d_sd[1].item() * 100, 1), "&",
np.round(dice_3d_mean[2].item() * 100, 1), " $\pm$ ",
np.round(dice_3d_sd[2].item() * 100, 1), "&",
np.round(dice_3d_mean[3].item() * 100, 1), " $\pm$ ",
np.round(dice_3d_sd[3].item() * 100, 1), "&",
np.round(dice_3d_mean.mean().item() * 100, 1), " $\pm$ ",
np.round(dice_3d_sd.mean().item() * 100, 1))
print("HD", np.round(hd_3d_mean[0].item(), 1), " $\pm$ ",
np.round(hd_3d_sd[0].item(), 1), "&",
np.round(hd_3d_mean[1].item(), 1), " $\pm$ ",
np.round(hd_3d_sd[1].item(), 1), "&",
np.round(hd_3d_mean[2].item(), 1), " $\pm$ ",
np.round(hd_3d_sd[2].item(), 1), "&",
np.round(hd_3d_mean[3].item(), 1), " $\pm$ ",
np.round(hd_3d_sd[3].item(), 1), "&",
np.round(hd_3d_mean.mean().item(), 1), " $\pm$ ",
np.round(hd_3d_sd.mean().item(), 1))
print("ASD", np.round(asd_3d_mean[0].item(), 1), " $\pm$ ",
np.round(asd_3d_sd[0].item(), 1), "&",
np.round(asd_3d_mean[1].item(), 1), " $\pm$ ",
np.round(asd_3d_sd[1].item(), 1), "&",
np.round(asd_3d_mean[2].item(), 1), " $\pm$ ",
np.round(asd_3d_sd[2].item(), 1), "&",
np.round(asd_3d_mean[3].item(), 1), " $\pm$ ",
np.round(asd_3d_sd[3].item(), 1), "&",
np.round(asd_3d_mean.mean().item(), 1), " $\pm$ ",
np.round(asd_3d_sd.mean().item(), 1))
[dice_3d, dice_3d_sd] = map_(lambda t: t.mean(),
[dice_3d_mean, dice_3d_sd])
[hd_3d, hd_3d_sd] = map_(lambda t: t.mean(), [hd_3d_mean, hd_3d_sd])
[asd_3d, asd_3d_sd] = map_(lambda t: t.mean(), [asd_3d_mean, asd_3d_sd])
if pprint:
print('asd_3d_mean', asd_3d_mean, "asd_3d_sd", asd_3d_sd, "hd_3d_mean",
hd_3d_mean, "hd_3d_sd", hd_3d_sd)
#return dice_3d.item(), dice_3d_sd.item(), hd_3d_mean.item(), hd_3d_sd.item(), asd_3d_mean.item(),asd_3d_sd.item()
return dice_3d.item(), dice_3d_sd.item(), hd_3d.item(), hd_3d_sd.item()
