# -*- coding: utf-8 -*-
import cv2
import numpy as np
import pandas as pd
def get_days(datetime):
return datetime.days
def get_sex(patient_sex):
return 1 if patient_sex == 'M' else 0
def min_max_scalar(x, min, max):
return (x - min) / (max - min)
def crop_volume(volume, crop):
volume_ = volume.copy()
volume_[:crop[0]] = 0
volume_[-crop[1]:] = 0
return volume_
def prob2binary(prob, thresh=0.5):
res = np.zeros_like(prob)
res[prob > thresh] = 1
return res
def resize(data, shape):
mask = np.zeros(shape)
for i in range(shape[0]):
mask[i, :, :] = cv2.resize(data[i, :, :], (shape[1], shape[2]))
return mask
def get_z(lesion):
z = 0
for i in range(lesion.shape[0]):
z = z + i * np.sum(lesion[i, :, :])
return z / (np.sum(lesion) + 1e-5)
def get_left_right(data, mid):
right = data[:, :, :mid]
left = data[:, :, mid:mid * 2]
return left, right
def get_consolidation(raw, lung, lesion, thresh=0.5):
""" Consolidation from raw, lung and lesion
"""
lung = prob2binary(lung)
lesion = prob2binary(lesion)
lung_lesion_union = lesion.astype(np.uint8) | lung.astype(np.uint8)
lung_lesion_union = lung_lesion_union.astype(np.uint8)
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))
lung_lesion_union_close = cv2.morphologyEx(
lung_lesion_union, cv2.MORPH_CLOSE, kernel)
kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (3, 3))
lung_lesion_union_open = cv2.morphologyEx(
lung_lesion_union_close, cv2.MORPH_CLOSE, kernel)
lung_lesion_union_open_area = raw * lung_lesion_union_open[:raw.shape[0]]
_, thres_image = cv2.threshold(
lung_lesion_union_open_area, thresh, 1, cv2.THRESH_BINARY)
kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (5, 5))
thres_image = cv2.morphologyEx(thres_image, cv2.MORPH_CLOSE, kernel)
kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (5, 1))
thres_image_open = cv2.morphologyEx(thres_image, cv2.MORPH_OPEN, kernel)
return thres_image_open
def calculate_volume(raw, lung, lesion, meta, crop=None):
if crop is None:
crop = [0.0, 0.0]
lung = prob2binary(lung)
lesion = prob2binary(lesion)
lung_lesion = lesion * lung
former_slice = 0
res_list = []
for index, row in meta.iterrows():
slices = row['slice']
spacing = eval(row['spacing'])
origin_shape = eval(row['shape'])
size_factor = (origin_shape[1] * origin_shape[2]
) / (lung.shape[1] * lung.shape[2])
voxel_size = spacing[0] * spacing[1] * spacing[2] * size_factor
total_slice, height, width = lung.shape
mid = int(width / 2)
current_slice = np.min([former_slice + slices, total_slice])
lung_current = lung[former_slice:current_slice]
lesion_current = lesion[former_slice:current_slice]
lung_lesion_current = lung_lesion[former_slice:current_slice]
raw_current = raw[former_slice:current_slice]
if crop[0] > 0:
lung_current = crop_volume(
lung_current, (np.array(crop) * slices).astype('int'))
lesion_current = crop_volume(
lesion_current, (np.array(crop) * slices).astype('int'))
lung_lesion_current = crop_volume(
lung_lesion_current, (np.array(crop) * slices).astype('int'))
left_lung, right_lung = get_left_right(lung_current, mid)
left_lesion, right_lesion = get_left_right(lung_lesion_current, mid)
left_raw, right_raw = get_left_right(raw_current, mid)
consolidation = get_consolidation(
raw_current, lung_current, lesion_current)
lesion_consolidation = lung_lesion_current * consolidation
left_consolidation, right_consolidation = get_left_right(
lesion_consolidation, mid)
calculate_list = [
lung_current, lesion_current, lung_lesion_current,
left_lung, right_lung, left_lesion, right_lesion,
lesion_current * raw_current, lung_lesion_current * raw_current,
left_lesion * left_raw, right_lesion * right_raw,
consolidation, lesion_consolidation,
left_consolidation, right_consolidation
]
[
lung_volume, lesion_volume, lung_lesion_volume,
left_lung_volume, right_lung_volume, left_lesion_volume, right_lesion_volume,
weighted_lesion_volume, weighted_lung_lesion_volume,
left_weighted_lesion_volume, right_weighted_lesion_volume,
consolidation_volume, lesion_consolidation_volume,
left_consolidation_volume, right_consolidation_volume
] = map(lambda x: np.sum(x) * voxel_size, calculate_list)
z = get_z(lung_lesion_current)
left_z = get_z(left_lesion)
right_z = get_z(right_lesion)
ratio = lung_lesion_volume / lung_volume
left_ratio = left_lesion_volume / left_lung_volume
right_ratio = right_lesion_volume / right_lung_volume
res_list.append(
{
'lung': lung_volume,
'lesion': lesion_volume,
'ratio': ratio,
'lung_lesion': lung_lesion_volume,
'left_lung': left_lung_volume,
'right_lung': right_lung_volume,
'left_lesion': left_lesion_volume,
'right_lesion': right_lesion_volume,
'left_ratio': left_ratio,
'right_ratio': right_ratio,
'weighted_lesion': weighted_lesion_volume,
'weighted_lung_lesion': weighted_lung_lesion_volume,
'left_weighted_lesion': left_weighted_lesion_volume,
'right_weighted_lesion': right_weighted_lesion_volume,
'consolidation': consolidation_volume,
'lesion_consolidation': lesion_consolidation_volume,
'left_consolidation': left_consolidation_volume,
'right_consolidation': right_consolidation_volume,
'z': z,
'left_z': left_z,
'right_z': right_z,
}
)
former_slice = current_slice
return res_list
def calculate(raw, lung, lesion, meta):
meta = meta[meta['slice'] > 300] # select thin scans
meta = meta.reset_index() # DataFrame index reset
res_list = calculate_volume(raw, lung, lesion, meta, crop=[0.17, 0.08])
all_info = pd.concat([meta, pd.DataFrame(res_list)], axis=1)
return res_list, all_info
def preprocessing(all_info, feature):
# transfer PatientSex into 0/1
all_info['sex'] = all_info['PatientSex'].map(get_sex)
# normalize the z-position by dividing the slice number of the CT scan
all_info['z'] = all_info['z'] / all_info['slice']
all_info['left_z'] = all_info['left_z'] / all_info['slice']
all_info['right_z'] = all_info['right_z'] / all_info['slice']
X = all_info[feature].astype(np.float32)
return X
Datasets
Models
