forked from scotthlee/hamlet
-
Notifications
You must be signed in to change notification settings - Fork 0
/
dataset_splitting_exp.py
315 lines (275 loc) · 12.3 KB
/
dataset_splitting_exp.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
import numpy as np
import pandas as pd
import os
from sklearn.model_selection import train_test_split
from hamlet.tools.generic import check_fnames, trim_zeroes
SEED = 2022
MIN_N = 1000
ADD_NEW_ONLY = False
# Setting the directorires
base_dir = 'C:/Users/yle4/data/hamlet/'
ds_dir = base_dir + 'misc/labels/'
if ADD_NEW_ONLY:
new_dir = base_dir + 'source/new/'
new_files = os.listdir(new_dir)
all_df = pd.read_csv(base_dir + 'all.csv')
samp_df = pd.read_csv(base_dir + 'samp.csv')
new_ids = [f[:-4] for f in new_files]
ids, args1, args2 = np.intersect1d(all_df.id.values,
new_ids,
return_indices=True)
good_df = all_df.loc[all_df.id.isin(ids)]
good_df['split'] = 'train'
good_df['file'] = [new_files[i] for i in args2]
good_df['batch'] = samp_df.batch.max() + 1
samp_df = pd.concat([samp_df, good_df], axis=0)
samp_df.to_csv(base_dir + 'samp.csv', index=False)
[os.rename(new_dir + f, base_dir + 'train/img/' + f)
for f in [new_files[i] for i in args2]]
# Reading the source Excel files
panels = [pd.read_csv(ds_dir + 'panel/' + f, encoding='latin')
for f in os.listdir(ds_dir + 'panel/')]
for df in panels:
ids = ['pan_' + s for s in df.ID.values.astype('str')]
df.ID = ids
immigrant = pd.read_csv(ds_dir + 'immigrant.csv', encoding='latin')
immigrant.ID = pd.Series(['im_' + s for s in immigrant.ID.astype('str')])
refugee = pd.read_csv(ds_dir + 'refugee.csv', encoding='latin')
refugee.ID = ['ref_' + s for s in refugee.ID.astype('str')]
non_iom = panels + [immigrant, refugee]
iom = pd.read_csv(ds_dir + 'iom.csv')
iom['sex'] = 'na'
# Renaming columns
col_dict = {'ID': 'id',
'panel_sitecode': 'panel_site',
'abnormal_img': 'abnormal',
'DS_ChestXrayFinding': 'abnormal',
'DS_Infiltrate': 'infiltrate',
'DS_ReticularMarkSuggestFibrosis': 'reticular',
'DS_CavitaryLesion': 'cavity',
'DS_Nodule': 'nodule',
'DS_Pleural': 'pleural_effusion',
'DS_HilarAdenopathy': 'hilar_adenopathy',
'DS_MiliaryFindings': 'miliary',
'DS_DiscreteLinearOpacity': 'linear_opacity',
'DS_DiscreteFibroticScar': 'linear_opacity',
'DS_DiscreteNodule': 'discrete_nodule',
'DS_DiscreteFibroticScarVolumeLoss': 'volume_loss',
'DS_IrregularThickPleuralReaction': 'pleural_reaction',
'DS_Other': 'other',
'DS_SCResults_Smear1Results': 'smear_1',
'DS_SCResults_Smear2Results': 'smear_2',
'DS_SCResults_Smear3Results': 'smear_3',
'DS_SCResults_Culture1Results': 'culture_1',
'DS_SCResults_Culture2Results': 'culture_2',
'DS_SCResults_Culture3Results': 'culture_3',
'DS_TBClass_NoClass': 'no_class',
'DS_TBClass_ClassA': 'class_a',
'DS_TBClass_ClassB1Pul': 'class_b1_pulm',
'DS_TBClass_ClassB1Extrapul': 'class_b1_extrapulm',
'DS_TBClass_ClassB2LTBI': 'class_b2_ltbi',
'DS_TBClass_ClassB3Contact': 'class_b3_contact',
'DS_TBClass_ClassBOther': 'class_b_other'
}
for df in non_iom:
df.columns = df.columns.str.replace(' ', '')
df.rename(columns=col_dict, inplace=True)
# Pulling out columns to make a combined dataset
abn_col = ['abnormal']
demo_cols = [
'id', 'exam_country', 'exam_date', 'birth_country',
'date_of_birth', 'panel_site', 'sex'
]
find_cols = [
'infiltrate', 'reticular', 'cavity',
'nodule', 'pleural_effusion', 'hilar_adenopathy',
'miliary', 'linear_opacity', 'discrete_nodule',
'volume_loss', 'pleural_reaction', 'other',
]
test_cols = [
'smear_1', 'smear_2', 'smear_3',
'culture_1', 'culture_2', 'culture_3'
]
class_cols = [
'class_a', 'class_b1_pulm', 'class_b1_extrapulm',
'class_b2_ltbi', 'class_b3_contact', 'class_b_other'
]
all_cols = demo_cols + abn_col + find_cols + test_cols + class_cols
# Merging the datasets
non_iom = [df[all_cols] for df in non_iom]
non_iom = [df.iloc[:, ~df.columns.duplicated()] for df in non_iom]
non_iom = pd.concat(non_iom, axis=0)
all_df = pd.concat([iom[all_cols], non_iom], axis=0)
all_df['abnormal_tb'] = np.array(all_df[find_cols].sum(axis=1) > 0,
dtype=np.uint8)
# Dropping rows without CXR readings
all_df = all_df.dropna(axis=0, subset=['abnormal'])
all_df = all_df.drop_duplicates(subset='id', keep='first')
# Making a data source variable
all_df['source'] = ''
all_df.source[['im_' in str(id) for id in all_df.id.values]] = 'immigrant'
all_df.source[['ref_' in str(id) for id in all_df.id.values]] = 'refugee'
all_df.source[['iom_' in str(id) for id in all_df.id.values]] = 'iom'
all_df.source[['pan_' in str(id) for id in all_df.id.values]] = 'panel'
# Making the first data flow table
n_tabs = pd.crosstab(all_df.source, 'n')
ab_tabs = pd.crosstab(all_df.source, all_df.abnormal)
ab_tabs['pct_ab'] = ab_tabs[1] / n_tabs.n
abtb_tabs = pd.crosstab(all_df.source, all_df.abnormal_tb)
abtb_tabs['pct_abtb'] = abtb_tabs[1] / n_tabs.n
tabs = pd.concat([n_tabs, ab_tabs, abtb_tabs], axis=1)
tabs.drop([0.0], axis=1, inplace=True)
tabs.columns = ['n', 'ab', 'pct_ab', 'abtb', 'pct_abtb']
tabs.to_csv(base_dir + 'all_ages_tab.csv')
# Dropping data from entrants under 15 years old
all_df['exam_date'] = pd.to_datetime(all_df.exam_date, errors='coerce')
all_df['date_of_birth'] = pd.to_datetime(all_df.date_of_birth, errors='coerce')
all_df.dropna(axis=0, inplace=True, subset=['exam_date', 'date_of_birth'])
ages = all_df.exam_date - all_df.date_of_birth
days = ages.dt.days.values
all_df['age_days'] = days
adults = np.where(days >= 15*365)[0]
kids = np.where(days < 15*365)[0]
all_df.iloc[kids, :].to_csv(base_dir + 'kids.csv', index=False)
all_df = all_df.iloc[adults, :].reset_index(drop=True)
# Making the source tables for adults
n_tabs = pd.crosstab(all_df.source, 'n')
ab_tabs = pd.crosstab(all_df.source, all_df.abnormal)
ab_tabs['pct_ab'] = ab_tabs[1] / n_tabs.n
abtb_tabs = pd.crosstab(all_df.source, all_df.abnormal_tb)
abtb_tabs['pct_abtb'] = abtb_tabs[1] / n_tabs.n
tabs = pd.concat([n_tabs, ab_tabs, abtb_tabs], axis=1)
tabs.drop([0.0], axis=1, inplace=True)
tabs.columns = ['n', 'ab', 'pct_ab', 'abtb', 'pct_abtb']
tabs.to_csv(base_dir + 'adults_tab.csv')
# Saving the dataset to file
all_df.to_csv(base_dir + 'all.csv', index=False)
presplit_dir = base_dir + 'presplit/'
fnames = os.listdir(presplit_dir)
short_fnames = [s[:-4] for s in fnames]
fname_dict = dict(zip(short_fnames, fnames))
# Quick check for images with no record
ids = all_df.id.values.astype('str')
no_record = np.setdiff1d(short_fnames, ids)
[os.rename(presplit_dir + fname_dict[f],
base_dir + 'source/bad/no_record/' + f[i])
for f in no_record]
fnames = os.listdir(presplit_dir)
short_fnames = [s[:-4] for s in fnames]
# Building the splits
has_img = np.intersect1d(all_df.id.values, short_fnames, return_indices=True)
fnames = [fname_dict[f] for f in has_img[0]]
short_fnames = [f[:-4] for f in fnames]
samp_df = all_df.iloc[has_img[1], :].drop_duplicates(subset='id')
samp_df = samp_df.reset_index(drop=True)
ids = samp_df.id.values.astype('str')
samp_df['file'] = [fname_dict[id] for id in ids]
# Making the flow table for adults with valid images
n_tabs = pd.crosstab(samp_df.source, 'n')
ab_tabs = pd.crosstab(samp_df.source, samp_df.abnormal)
ab_tabs['pct_ab'] = ab_tabs[1] / n_tabs.n
abtb_tabs = pd.crosstab(samp_df.source, samp_df.abnormal_tb)
abtb_tabs['pct_abtb'] = abtb_tabs[1] / n_tabs.n
tabs = pd.concat([n_tabs, ab_tabs, abtb_tabs], axis=1)
tabs.drop([0.0], axis=1, inplace=True)
tabs.columns = ['n', 'ab', 'pct_ab', 'abtb', 'pct_abtb']
tabs.to_csv(base_dir + 'valid_tab.csv')
# Specifying the panel sites to use for reference reads
sites = samp_df.panel_site.values.astype('str')
good_sites = [
'Cho Ray', 'ASVIET1', 'Luke',
'ASPHIL1', 'Consultorios de Visa', 'AMDOMI1',
'AMDOMI2', 'Servicios Medicos Consulares', 'AMMEXI1',
'Clinica Medical Internacional', 'AMMEXI2',
'Medicos Especializados', 'AMMEXI3',
'Servicios Medicos de la Frontera'
]
good_any = np.array([s in good_sites for s in sites], dtype=np.uint8)
panel_pos = np.array(['pan_' in s for s in ids], dtype=np.uint8)
iom_read = np.array(['iom_' in s for s in ids], dtype=np.uint8)
pref_reads = np.array(good_any + panel_pos + iom_read > 0, dtype=np.uint8)
# Setting up the reference reads for validation and testing
abtb = samp_df.abnormal_tb.values
ab = samp_df.abnormal.values
abnotb = np.array((ab == 1) & (abtb == 0), dtype=np.uint8)
samp_df['ab_kind'] = 'normal'
samp_df.ab_kind[abnotb == 1] = 'no TB'
samp_df.ab_kind[abtb == 1] = 'TB'
# Setting up the sampling ratios for two settings: high TB and low TB;
# probabilities are for TB, no TB, and normal; these probabilities are about
# the same as they are in all_df; NOTE this splitting technique may introduce
# substantial noise into the validation and test datasets from the non-
# preferred-read abnormal images.
source = samp_df.source.values.astype(str)
high_tb_cond = (good_any == 1) & ([s in ['immigrant', 'refugee']
for s in source])
high_tb = pd.crosstab(samp_df.ab_kind[high_tb_cond], 'n')
high_tb_p = (high_tb / high_tb.sum()).values.flatten()
low_tb_cond = (good_any == 0) & ([s in ['immigrant', 'refugee']
for s in source])
low_tb = pd.crosstab(samp_df.ab_kind[low_tb_cond], 'n')
low_tb_p = (low_tb / low_tb.sum()).values.flatten()
mean_p = np.mean([high_tb_p, low_tb_p], axis=0)
# Calculating how many images (in total) to use for validation and testing
# for each of the two scenarios
scenarios = [high_tb_p, low_tb_p]
prob_names = ['TB', 'no TB', 'normal']
scen_counts = []
for i, s in enumerate(scenarios):
p_sort = np.argsort(s[0:2])
ratio = s[p_sort[1]] / s[p_sort[0]]
other_n = int(ratio * MIN_N)
p_names = [prob_names[j] for j in p_sort]
scen_counts.append(dict(zip(p_names, [MIN_N * 2, other_n * 2])))
# Building the samples
np.random.seed(SEED)
scen_samps = []
ref_abtb = np.array((pref_reads == 1) & (abtb == 1))
ref_notb = np.array(abnotb == 1)
ref_norm = np.array((pref_reads == 1) & (ab == 0))
total_tb = np.sum([s['TB'] for s in scen_counts])
total_notb = np.sum([s['no TB'] for s in scen_counts])
total_norm = MIN_N * 4
tb_ids = np.random.choice(ids[ref_abtb], size=total_tb, replace=False)
notb_ids = np.random.choice(ids[ref_notb], size=total_notb, replace=False)
norm_ids = np.random.choice(ids[ref_norm], size=total_norm, replace=False)
high_tb_ids = [
np.random.choice(tb_ids, size=scen_counts[0]['TB'], replace=False),
np.random.choice(notb_ids, size=scen_counts[0]['no TB'], replace=False),
np.random.choice(norm_ids, size=MIN_N * 2, replace=False)
]
low_tb_ids = [
np.setdiff1d(tb_ids, high_tb_ids[0]),
np.setdiff1d(notb_ids, high_tb_ids[1]),
np.setdiff1d(norm_ids, high_tb_ids[2])
]
high_tb_ids = np.concatenate(high_tb_ids)
low_tb_ids = np.concatenate(low_tb_ids)
val_high = np.random.choice(high_tb_ids,
size=int(.5 * len(high_tb_ids)),
replace=False)
val_low = np.random.choice(low_tb_ids,
size=int(.5 * len(low_tb_ids)),
replace=False)
test_high = np.setdiff1d(high_tb_ids, val_high)
test_low = np.setdiff1d(low_tb_ids, val_low)
# Assigning everythign else to training data and then smushing together the
# validation IDs (keeping them separate is only optional)
train_ids = np.setdiff1d(ids, np.concatenate([high_tb_ids,
low_tb_ids]).flatten())
val_ids = np.concatenate([val_high, val_low]).flatten()
# Making a lookup dictionary that specifies the split
id_dict = dict(zip(val_ids, ['val'] * len(val_ids)))
id_dict.update(dict(zip(train_ids, ['train'] * len(train_ids))))
id_dict.update(dict(zip(test_high, ['test_high'] * len(test_high))))
id_dict.update(dict(zip(test_low, ['test_low'] * len(test_low))))
# Writing the CSV back to disk with the split info
samp_df['split'] = [id_dict[id] for id in samp_df.id]
samp_df.to_csv(base_dir + 'samp.csv', index=False)
# And now moving the validation and test images
split_dict = dict(zip(samp_df.id.values,
samp_df.split.values))
for f in samp_df.file.values:
ds = split_dict[f[:-4]]
path = base_dir + ds + '/img/'
os.rename(presplit_dir + f, path + f)