161 lines
6.0 KiB
Python
Executable File
161 lines
6.0 KiB
Python
Executable File
import argparse
|
|
from os import path
|
|
import SimpleITK as sitk
|
|
import tensorflow as tf
|
|
from tensorflow.keras.models import load_model
|
|
import numpy as np
|
|
import os
|
|
|
|
from sfransen.utils_quintin import *
|
|
from sfransen.DWI_exp import preprocess
|
|
from sfransen.DWI_exp.helpers import *
|
|
from sfransen.DWI_exp.callbacks import dice_coef
|
|
from sfransen.DWI_exp.losses import weighted_binary_cross_entropy
|
|
|
|
from sfransen.FROC.blob_preprocess import *
|
|
from sfransen.FROC.cal_froc_from_np import *
|
|
from sfransen.load_images import load_images_parrallel
|
|
from sfransen.Saliency.base import *
|
|
from sfransen.Saliency.integrated_gradients import *
|
|
|
|
parser = argparse.ArgumentParser(
|
|
description='Calculate the froc metrics and store in froc_metrics.yml')
|
|
parser.add_argument('-experiment',
|
|
help='Title of experiment')
|
|
parser.add_argument('--series', '-s',
|
|
metavar='[series_name]', required=True, nargs='+',
|
|
help='List of series to include')
|
|
args = parser.parse_args()
|
|
|
|
def print_p(*args, **kwargs):
|
|
"""
|
|
Shorthand for print(..., flush=True)
|
|
Useful on HPC cluster where output has buffered writes.
|
|
"""
|
|
print(*args, **kwargs, flush=True)
|
|
|
|
######## CUDA ################
|
|
os.environ["CUDA_VISIBLE_DEVICES"] = "2"
|
|
|
|
######## constants #############
|
|
SERIES = args.series
|
|
series_ = '_'.join(args.series)
|
|
EXPERIMENT = args.experiment
|
|
|
|
DATA_DIR = "./../data/Nijmegen paths/"
|
|
TARGET_SPACING = (0.5, 0.5, 3)
|
|
INPUT_SHAPE = (192, 192, 24, len(SERIES))
|
|
IMAGE_SHAPE = INPUT_SHAPE[:3]
|
|
|
|
image_paths = {}
|
|
for s in SERIES:
|
|
with open(path.join(DATA_DIR, f"{s}.txt"), 'r') as f:
|
|
image_paths[s] = [l.strip() for l in f.readlines()]
|
|
with open(path.join(DATA_DIR, f"seg.txt"), 'r') as f:
|
|
seg_paths = [l.strip() for l in f.readlines()]
|
|
num_images = len(seg_paths)
|
|
|
|
max_saliency_values = []
|
|
for fold in range(5):
|
|
print_p("fold:",fold)
|
|
|
|
# model path
|
|
MODEL_PATH = f'./../train_output/{EXPERIMENT}_{series_}_{fold}/models/{EXPERIMENT}_{series_}_{fold}.h5'
|
|
YAML_DIR = f'./../train_output/{EXPERIMENT}_{series_}_{fold}'
|
|
IMAGE_DIR = f'./../train_output/{EXPERIMENT}_{series_}_{fold}'
|
|
|
|
# test indices
|
|
DATA_SPLIT_INDEX = read_yaml_to_dict(f'./../data/Nijmegen paths/train_val_test_idxs_{fold}.yml')
|
|
TEST_INDEX_IMGS = DATA_SPLIT_INDEX['test_set0']
|
|
|
|
for img_idx in TEST_INDEX_IMGS[:10]:
|
|
print_p("img_idx:",img_idx)
|
|
images = []
|
|
images_list = []
|
|
segmentations = []
|
|
saliency_map = []
|
|
|
|
# Read and preprocess each of the paths for each series, and the segmentations.
|
|
# print('images number',[TEST_INDEX[img_idx]])
|
|
img_s = {f'{s}': sitk.ReadImage(image_paths[s][img_idx], sitk.sitkFloat32) for s in SERIES}
|
|
seg_s = sitk.ReadImage(seg_paths[img_idx], sitk.sitkFloat32)
|
|
img_n, seg_n = preprocess(img_s, seg_s,
|
|
shape=IMAGE_SHAPE, spacing=TARGET_SPACING)
|
|
for seq in img_n:
|
|
images.append(img_n[f'{seq}'])
|
|
images_list.append(images)
|
|
images = []
|
|
segmentations.append(seg_n)
|
|
|
|
images_list = np.transpose(images_list, (0, 2, 3, 4, 1))
|
|
|
|
|
|
########### load module ##################
|
|
# print(' >>>>>>> LOAD MODEL <<<<<<<<<')
|
|
|
|
dependencies = {
|
|
'dice_coef': dice_coef,
|
|
'weighted_cross_entropy_fn':weighted_binary_cross_entropy
|
|
}
|
|
reconstructed_model = load_model(MODEL_PATH, custom_objects=dependencies)
|
|
# reconstructed_model.summary(line_length=120)
|
|
|
|
# make predictions on all TEST_INDEX
|
|
# print(' >>>>>>> START prediction <<<<<<<<<')
|
|
predictions_blur = reconstructed_model.predict(images_list, batch_size=1)
|
|
|
|
|
|
############# preprocess #################
|
|
# preprocess predictions by removing the blur and making individual blobs
|
|
# print('>>>>>>>> START preprocess')
|
|
|
|
def move_dims(arr):
|
|
# UMCG numpy dimensions convention: dims = (batch, width, heigth, depth)
|
|
# Joeran numpy dimensions convention: dims = (batch, depth, heigth, width)
|
|
arr = np.moveaxis(arr, 3, 1)
|
|
arr = np.moveaxis(arr, 3, 2)
|
|
return arr
|
|
|
|
# Joeran has his numpy arrays ordered differently.
|
|
|
|
predictions_blur = move_dims(np.squeeze(predictions_blur,axis=4))
|
|
segmentations = move_dims(segmentations)
|
|
# predictions = [preprocess_softmax(pred, threshold="dynamic")[0] for pred in predictions_blur]
|
|
predictions = predictions_blur
|
|
# print("the size of predictions is:",np.shape(predictions))
|
|
# Remove outer edges
|
|
zeros = np.zeros(np.shape(predictions))
|
|
test = predictions[:,2:-2,2:190,2:190]
|
|
zeros[:,2:-2,2:190,2:190] = test
|
|
predictions = zeros
|
|
# print(np.shape(predictions))
|
|
######### Build Saliency heatmap ##############
|
|
# print(' >>>>>>> Build saliency map <<<<<<<<<')
|
|
|
|
ig = IntegratedGradients(reconstructed_model)
|
|
for img_idx in range(len(images_list)):
|
|
# input_img = np.resize(images_list[img_idx],(1,192,192,24,len(SERIES)))
|
|
saliency_map.append(ig.get_mask(images_list).numpy())
|
|
|
|
# print("size saliency map",np.shape(saliency_map))
|
|
|
|
idx_max = np.argmax(np.mean(np.mean(np.mean(np.squeeze(saliency_map),axis=0),axis=0),axis=0))
|
|
|
|
max_saliency_values.append(idx_max)
|
|
print_p("max_saliency_values:",max_saliency_values)
|
|
|
|
t2_max = sum(map(lambda x : x == 0, max_saliency_values))
|
|
dwi_max = sum(map(lambda x : x == 1, max_saliency_values))
|
|
adc_max = sum(map(lambda x : x == 2, max_saliency_values))
|
|
|
|
print_p(f"max value in t2 is: {t2_max}")
|
|
print_p(f"max value in dwi is: {dwi_max}")
|
|
print_p(f"max value in adc is: {adc_max}")
|
|
|
|
# np.save(f'{YAML_DIR}/saliency_new23',saliency_map)
|
|
# np.save(f'{YAML_DIR}/images_list_new23',images_list)
|
|
# np.save(f'{YAML_DIR}/segmentations_new23',segmentations)
|
|
# np.save(f'{YAML_DIR}/predictions_new23',predictions)
|
|
|
|
|