def predict_RF(train_pats, test_pats, fscores=None, do_plot_predictions=False,
stratified=False):
"""
Predict tumor voxels for given test patients.
Input:
train_pats -- list of patient IDs used for training a model.
test_pats -- list of patient IDs used for testing a model.
fscores -- An opened output file to which we write the results.
"""
xtr, ytr, coordtr, patient_idxs_tr, dims_tr = dp.load_patients(train_pats,
stratified)
if stratified:
# Class frequencies in the whole dataset
class_freqs = dp.class_counts / float(sum(dp.class_counts))
print "Class frequencies:", class_freqs*100
# Class frequencies in the sample
sample_counts = np.histogram(ytr, range(6))[0]
sample_freqs = sample_counts / float(sum(sample_counts))
print "Sample frequencies:", sample_freqs*100
weights = np.ones(len(ytr))
for i in range(5):
weights[ytr==i] = class_freqs[i] / sample_freqs[i]
else:
weights = None
model = train_RF_model(xtr, ytr, n_trees=30, sample_weight=weights)
print "\n----------------------------------\n"
yte = np.zeros(0)
predte = np.zeros(0)
patient_idxs_te = [0]
print "Test users:"
# Iterate over test users
for te_idx, te_pat in enumerate(test_pats):
print "Test patient number %d" % (te_idx+1)
x, y, coord, dim = dp.load_patient(te_pat, n_voxels=None)
if do_plot_predictions:
pif = os.path.join('plots', 'pat%d_slices_0_RF.png' % te_pat)
else:
pif = None
pred = predict_and_evaluate(
model, x, y, coord=coord, dim_list=[dim], plot_confmat=False,
ret_probs=False, patient_idxs=None,
pred_img_fname=pif)
#pred = np.argmax(pred_probs_te, axis=1)
yte = np.concatenate((yte, y))
patient_idxs_te.append(len(yte))
predte = np.concatenate((predte, pred))
'''
for i in range(1):
xlabel_te = dp.extract_label_features(coordte, dims_te, pred_probs_te,
patient_idxs_te)
smoothed_pred = np.argmax(xlabel_te, axis=1)
dice_scores(yte, smoothed_pred, patient_idxs=patient_idxs_te,
label='Test smoothed dice scores (iteration %d):' % (i+1))
xte2 = np.hstack((xte, xlabel_te))
pred_probs_te = predict_and_evaluate(
model2, xte2, yte, coord=coordte, dim_list=dims_te, pred_fname=None,
plot_confmat=False, ret_probs=True, patient_idxs=patient_idxs_te,
pred_img_fname=os.path.join('plots', 'pat%d_slices_%d.png' % (test_pats[0], i+1)))
'''
print "\nOverall confusion matrix:"
cm = confusion_matrix(yte, predte)
print cm
dice_scores(yte, predte, patient_idxs=patient_idxs_te,
label='Overall dice scores (RF):', fscores=fscores)
def predict_RF(train_pats,
test_pats,
fscores=None,
do_plot_predictions=False,
stratified=False):
"""
Predict tumor voxels for given test patients.
Input:
train_pats -- list of patient IDs used for training a model.
test_pats -- list of patient IDs used for testing a model.
fscores -- An opened output file to which we write the results.
"""
xtr, ytr, coordtr, patient_idxs_tr, dims_tr = dp.load_patients(
train_pats, stratified)
if stratified:
# Class frequencies in the whole dataset
class_freqs = dp.class_counts / float(sum(dp.class_counts))
print "Class frequencies:", class_freqs * 100
# Class frequencies in the sample
sample_counts = np.histogram(ytr, range(6))[0]
sample_freqs = sample_counts / float(sum(sample_counts))
print "Sample frequencies:", sample_freqs * 100
weights = np.ones(len(ytr))
for i in range(5):
weights[ytr == i] = class_freqs[i] / sample_freqs[i]
else:
weights = None
model = train_RF_model(xtr, ytr, n_trees=30, sample_weight=weights)
print "\n----------------------------------\n"
yte = np.zeros(0)
predte = np.zeros(0)
patient_idxs_te = [0]
print "Test users:"
# Iterate over test users
for te_idx, te_pat in enumerate(test_pats):
print "Test patient number %d" % (te_idx + 1)
x, y, coord, dim = dp.load_patient(te_pat, n_voxels=None)
if do_plot_predictions:
pif = os.path.join('plots', 'pat%d_slices_0_RF.png' % te_pat)
else:
pif = None
pred = predict_and_evaluate(model,
x,
y,
coord=coord,
dim_list=[dim],
plot_confmat=False,
ret_probs=False,
patient_idxs=None,
pred_img_fname=pif)
#pred = np.argmax(pred_probs_te, axis=1)
yte = np.concatenate((yte, y))
patient_idxs_te.append(len(yte))
predte = np.concatenate((predte, pred))
'''
for i in range(1):
xlabel_te = dp.extract_label_features(coordte, dims_te, pred_probs_te,
patient_idxs_te)
smoothed_pred = np.argmax(xlabel_te, axis=1)
dice_scores(yte, smoothed_pred, patient_idxs=patient_idxs_te,
label='Test smoothed dice scores (iteration %d):' % (i+1))
xte2 = np.hstack((xte, xlabel_te))
pred_probs_te = predict_and_evaluate(
model2, xte2, yte, coord=coordte, dim_list=dims_te, pred_fname=None,
plot_confmat=False, ret_probs=True, patient_idxs=patient_idxs_te,
pred_img_fname=os.path.join('plots', 'pat%d_slices_%d.png' % (test_pats[0], i+1)))
'''
print "\nOverall confusion matrix:"
cm = confusion_matrix(yte, predte)
print cm
dice_scores(yte,
predte,
patient_idxs=patient_idxs_te,
label='Overall dice scores (RF):',
fscores=fscores)
def predict_two_stage(train_pats, test_pats, fscores=None,
do_plot_predictions=False, stratified=False, n_trees=30,
dev_pats=[], use_mrf=True, resolution=1, n_voxels=30000,
mat_dir=None, fresh_models=False, load_hog=False):
"""
Predict tumor voxels for given test patients.
Input:
train_pats -- list of patient IDs used for training a model.
test_pats -- list of patient IDs used for testing a model.
fscores -- An opened output file to which we write the results.
"""
model_str = ""
if resolution != 1:
model_str += '_res%d' % resolution
if load_hog:
model_str += '_hog'
model1_fname = os.path.join('models', 'model1_seed%d_ntrp%d_ntep%d_ntrees%d_nvox%s%s.jl' %
(seed, len(train_pats), len(test_pats), n_trees, n_voxels, model_str))
model2_fname = os.path.join('models', 'model2_seed%d_ntrp%d_ntep%d_ntrees%d_nvox%s%s.jl' %
(seed, len(train_pats), len(test_pats), n_trees, n_voxels, model_str))
# Load models if available
if not fresh_models and os.path.isfile(model1_fname) and \
os.path.isfile(model2_fname):
model1 = joblib.load(model1_fname)
model2 = joblib.load(model2_fname)
min_voxels = 3000
else:
xtr, ytr, coordtr, patient_idxs_tr, dims_tr = dp.load_patients(
train_pats, stratified, resolution=resolution,
n_voxels=n_voxels, load_hog=load_hog)
# Make all tumor labels equal to 1 and train the first model
ytr1 = np.array(ytr, copy=True)
ytr1[ytr1>0] = 1
if stratified:
# Class frequencies in the whole dataset
class_counts = [dp.class_counts[0], sum(dp.class_counts[1:])]
class_freqs = np.asarray(class_counts) / float(sum(class_counts))
print "Class frequencies (model 1):", class_freqs*100
# Class frequencies in the sample
sample_counts = np.histogram(ytr, [0,1,5])[0]
sample_freqs = sample_counts / float(sum(sample_counts))
print "Sample frequencies:", sample_freqs*100
weights = np.ones(len(ytr))
for i in range(2):
weights[ytr==i] = class_freqs[i] / sample_freqs[i]
else:
weights = None
model1 = train_RF_model(xtr, ytr1, n_trees=n_trees,
sample_weight=weights, fname=model1_fname)
# Compute minimum number of tumor voxels in a train patient
min_voxels = 3000#get_min_voxels(ytr, patient_idxs_tr)
print "Minimum number of voxels in a tumor: %d" % min_voxels
# Train the second model to separate tumor classes
ok_idxs = ytr > 0
xtr2 = np.asarray(xtr[ok_idxs,:])
ytr2 = np.asarray(ytr[ok_idxs])
if stratified:
# Class frequencies in the whole dataset
class_counts = dp.class_counts[1:]
class_freqs = np.asarray(class_counts) / float(sum(class_counts))
print "Class frequencies (model 2):", class_freqs*100
# Class frequencies in the sample
sample_counts = np.histogram(ytr, range(1,6))[0]
sample_freqs = sample_counts / float(sum(sample_counts))
print "Sample frequencies:", sample_freqs*100
weights = np.ones(len(ytr2))
for i in range(4):
weights[ytr2==i+1] = class_freqs[i] / sample_freqs[i]
else:
weights = None
model2 = train_RF_model(xtr2, ytr2, n_trees=n_trees,
sample_weight=weights, fname=model2_fname)
print "\n----------------------------------\n"
if len(dev_pats) > 0:
best_potential = optimize_potential(
dev_pats, model1, model2, stratified, fscores,
do_plot_predictions, resolution=resolution, load_hog=load_hog)
best_radius = optimize_closing(dev_pats, model1, stratified, fscores,
resolution=resolution, load_hog=load_hog)
best_th = optimize_threshold1(dev_pats, model1, stratified, fscores,
resolution, load_hog, best_radius)
else:
best_radius = 6
best_th = 0.6
best_potential = np.array([[0.04, 0.03555556, 0.03555556, 0.02222222],
[0.03555556, 0.04, 0.02222222, 0.],
[0.03555556, 0.02222222, 0.04, 0.03555556],
[0.02222222, 0., 0.03555556, 0.04]])
yte = np.zeros(0)
predte = np.zeros(0)
predte_no_pp = np.zeros(0)
patient_idxs_te = [0]
print "Test users:"
# Iterate over test users
for te_idx, te_pat in enumerate(test_pats):
print "Test patient number %d" % (te_idx+1)
x, y, coord, dim = dp.load_patient(te_pat, n_voxels=None,
resolution=resolution,
load_hog=load_hog)
#pred = model1.predict(x)
pred_probs = model1.predict_proba(x)
#pred = np.argmax(pred_probs, axis=1)
pred = pred_probs[:,1] >= best_th
# If the predicted tumor is too small set the most probable tumor
# voxels to one
if sum(pred > 0) < min_voxels:
print "Patient having too few voxels (%d < %d)" % (sum(pred > 0), min_voxels)
pred = np.zeros(pred.shape)
new_idxs = np.argsort(pred_probs[:,1])[-min_voxels:]
pred[new_idxs] = 1
pp_pred = dp.post_process(coord, dim, pred, binary_closing=True,
radius=best_radius)
tumor_idxs = pp_pred > 0
if sum(tumor_idxs) > 0:
pred_probs2 = model2.predict_proba(x[tumor_idxs,:])
pred2 = np.argmax(pred_probs2, axis=1) + 1
pp_pred[tumor_idxs] = pred2
pp_pred15 = np.array(pp_pred)
print "\nConfusion matrix:"
cm = confusion_matrix(y, pp_pred15)
print cm
dice_scores(y, pp_pred15, label='Dice scores:')
if use_mrf:
# MRF post processing
if sum(tumor_idxs) > 0:
edges = dp.create_graph(coord[tumor_idxs,:])
pp_pred[tumor_idxs] = dp.mrf(pred_probs2, edges,
potential=best_potential) + 1
method = 'MRF'
else:
# Closing post processing
if sum(tumor_idxs) > 0:
pp_pred[tumor_idxs] = dp.post_process(coord[tumor_idxs,:], dim,
pred2, remove_components=False,
radius=best_radius)
method = 'closing'
print "\nConfusion matrix (pp):"
cm = confusion_matrix(y, pp_pred)
print cm
yte = np.concatenate((yte, y))
patient_idxs_te.append(len(yte))
predte = np.concatenate((predte, pp_pred))
predte_no_pp = np.concatenate((predte_no_pp, pp_pred15))
dice_scores(y, pp_pred, label='Dice scores (pp):')
if do_plot_predictions:
# Plot the patient
pif = os.path.join('results', 'pat%d_slices_2S_%s.png' % (te_pat, method))
if mat_dir is not None:
fmat = os.path.join(mat_dir, 'pat%d.mat' % te_pat)
else:
fmat = None
pp.plot_predictions(coord, dim, pp_pred15, y, pp_pred, fname=pif,
fmat=fmat)
#if pred_fname is not None:
# extras.save_predictions(coord, dim_list[0], pred, yte, pred_fname)
print "\nOverall confusion matrix:"
cm = confusion_matrix(yte, predte)
print cm
dice_scores(yte, predte_no_pp, patient_idxs=patient_idxs_te,
label='Overall dice scores (two-stage, no pp):', fscores=fscores)
dice_scores(yte, predte, patient_idxs=patient_idxs_te,
label='Overall dice scores (two-stage):', fscores=fscores)
def predict_two_stage(train_pats,
test_pats,
fscores=None,
do_plot_predictions=False,
stratified=False,
n_trees=30,
dev_pats=[],
use_mrf=True,
resolution=1,
n_voxels=30000,
mat_dir=None,
fresh_models=False,
load_hog=False):
"""
Predict tumor voxels for given test patients.
Input:
train_pats -- list of patient IDs used for training a model.
test_pats -- list of patient IDs used for testing a model.
fscores -- An opened output file to which we write the results.
"""
model_str = ""
if resolution != 1:
model_str += '_res%d' % resolution
if load_hog:
model_str += '_hog'
model1_fname = os.path.join(
'models', 'model1_seed%d_ntrp%d_ntep%d_ntrees%d_nvox%s%s.jl' %
(seed, len(train_pats), len(test_pats), n_trees, n_voxels, model_str))
model2_fname = os.path.join(
'models', 'model2_seed%d_ntrp%d_ntep%d_ntrees%d_nvox%s%s.jl' %
(seed, len(train_pats), len(test_pats), n_trees, n_voxels, model_str))
# Load models if available
if not fresh_models and os.path.isfile(model1_fname) and \
os.path.isfile(model2_fname):
model1 = joblib.load(model1_fname)
model2 = joblib.load(model2_fname)
min_voxels = 3000
else:
xtr, ytr, coordtr, patient_idxs_tr, dims_tr = dp.load_patients(
train_pats,
stratified,
resolution=resolution,
n_voxels=n_voxels,
load_hog=load_hog)
# Make all tumor labels equal to 1 and train the first model
ytr1 = np.array(ytr, copy=True)
ytr1[ytr1 > 0] = 1
if stratified:
# Class frequencies in the whole dataset
class_counts = [dp.class_counts[0], sum(dp.class_counts[1:])]
class_freqs = np.asarray(class_counts) / float(sum(class_counts))
print "Class frequencies (model 1):", class_freqs * 100
# Class frequencies in the sample
sample_counts = np.histogram(ytr, [0, 1, 5])[0]
sample_freqs = sample_counts / float(sum(sample_counts))
print "Sample frequencies:", sample_freqs * 100
weights = np.ones(len(ytr))
for i in range(2):
weights[ytr == i] = class_freqs[i] / sample_freqs[i]
else:
weights = None
model1 = train_RF_model(xtr,
ytr1,
n_trees=n_trees,
sample_weight=weights,
fname=model1_fname)
# Compute minimum number of tumor voxels in a train patient
min_voxels = 3000 #get_min_voxels(ytr, patient_idxs_tr)
print "Minimum number of voxels in a tumor: %d" % min_voxels
# Train the second model to separate tumor classes
ok_idxs = ytr > 0
xtr2 = np.asarray(xtr[ok_idxs, :])
ytr2 = np.asarray(ytr[ok_idxs])
if stratified:
# Class frequencies in the whole dataset
class_counts = dp.class_counts[1:]
class_freqs = np.asarray(class_counts) / float(sum(class_counts))
print "Class frequencies (model 2):", class_freqs * 100
# Class frequencies in the sample
sample_counts = np.histogram(ytr, range(1, 6))[0]
sample_freqs = sample_counts / float(sum(sample_counts))
print "Sample frequencies:", sample_freqs * 100
weights = np.ones(len(ytr2))
for i in range(4):
weights[ytr2 == i + 1] = class_freqs[i] / sample_freqs[i]
else:
weights = None
model2 = train_RF_model(xtr2,
ytr2,
n_trees=n_trees,
sample_weight=weights,
fname=model2_fname)
print "\n----------------------------------\n"
if len(dev_pats) > 0:
best_potential = optimize_potential(dev_pats,
model1,
model2,
stratified,
fscores,
do_plot_predictions,
resolution=resolution,
load_hog=load_hog)
best_radius = optimize_closing(dev_pats,
model1,
stratified,
fscores,
resolution=resolution,
load_hog=load_hog)
best_th = optimize_threshold1(dev_pats, model1, stratified, fscores,
resolution, load_hog, best_radius)
else:
best_radius = 6
best_th = 0.6
best_potential = np.array([[0.04, 0.03555556, 0.03555556, 0.02222222],
[0.03555556, 0.04, 0.02222222, 0.],
[0.03555556, 0.02222222, 0.04, 0.03555556],
[0.02222222, 0., 0.03555556, 0.04]])
yte = np.zeros(0)
predte = np.zeros(0)
predte_no_pp = np.zeros(0)
patient_idxs_te = [0]
print "Test users:"
# Iterate over test users
for te_idx, te_pat in enumerate(test_pats):
print "Test patient number %d" % (te_idx + 1)
x, y, coord, dim = dp.load_patient(te_pat,
n_voxels=None,
resolution=resolution,
load_hog=load_hog)
#pred = model1.predict(x)
pred_probs = model1.predict_proba(x)
#pred = np.argmax(pred_probs, axis=1)
pred = pred_probs[:, 1] >= best_th
# If the predicted tumor is too small set the most probable tumor
# voxels to one
if sum(pred > 0) < min_voxels:
print "Patient having too few voxels (%d < %d)" % (sum(pred > 0),
min_voxels)
pred = np.zeros(pred.shape)
new_idxs = np.argsort(pred_probs[:, 1])[-min_voxels:]
pred[new_idxs] = 1
pp_pred = dp.post_process(coord,
dim,
pred,
binary_closing=True,
radius=best_radius)
tumor_idxs = pp_pred > 0
if sum(tumor_idxs) > 0:
pred_probs2 = model2.predict_proba(x[tumor_idxs, :])
pred2 = np.argmax(pred_probs2, axis=1) + 1
pp_pred[tumor_idxs] = pred2
pp_pred15 = np.array(pp_pred)
print "\nConfusion matrix:"
cm = confusion_matrix(y, pp_pred15)
print cm
dice_scores(y, pp_pred15, label='Dice scores:')
if use_mrf:
# MRF post processing
if sum(tumor_idxs) > 0:
edges = dp.create_graph(coord[tumor_idxs, :])
pp_pred[tumor_idxs] = dp.mrf(
pred_probs2, edges, potential=best_potential) + 1
method = 'MRF'
else:
# Closing post processing
if sum(tumor_idxs) > 0:
pp_pred[tumor_idxs] = dp.post_process(coord[tumor_idxs, :],
dim,
pred2,
remove_components=False,
radius=best_radius)
method = 'closing'
print "\nConfusion matrix (pp):"
cm = confusion_matrix(y, pp_pred)
print cm
yte = np.concatenate((yte, y))
patient_idxs_te.append(len(yte))
predte = np.concatenate((predte, pp_pred))
predte_no_pp = np.concatenate((predte_no_pp, pp_pred15))
dice_scores(y, pp_pred, label='Dice scores (pp):')
if do_plot_predictions:
# Plot the patient
pif = os.path.join('results',
'pat%d_slices_2S_%s.png' % (te_pat, method))
if mat_dir is not None:
fmat = os.path.join(mat_dir, 'pat%d.mat' % te_pat)
else:
fmat = None
pp.plot_predictions(coord,
dim,
pp_pred15,
y,
pp_pred,
fname=pif,
fmat=fmat)
#if pred_fname is not None:
# extras.save_predictions(coord, dim_list[0], pred, yte, pred_fname)
print "\nOverall confusion matrix:"
cm = confusion_matrix(yte, predte)
print cm
dice_scores(yte,
predte_no_pp,
patient_idxs=patient_idxs_te,
label='Overall dice scores (two-stage, no pp):',
fscores=fscores)
dice_scores(yte,
predte,
patient_idxs=patient_idxs_te,
label='Overall dice scores (two-stage):',
fscores=fscores)
