-
Notifications
You must be signed in to change notification settings - Fork 1
/
fp_reducer.py
196 lines (144 loc) · 7.5 KB
/
fp_reducer.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
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Sat May 30 15:12:29 2020
@author: Yesh
"""
import os # os.chdir('/Users/Yesh/Documents/BDRAD/chest_ct_projects/pytorch-retinanet')
import numpy as np
import pandas as pd
import time
import argparse
import matplotlib.pyplot as plt
from xgboost import XGBClassifier
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import StratifiedKFold, GridSearchCV, train_test_split
from sklearn.metrics import accuracy_score, confusion_matrix, roc_auc_score, roc_curve, classification_report
from joblib import dump
plt.rcParams['figure.figsize'] = [12, 6]
"""
Train ML model to reduce false positives using clustering metadata
Ideas:
- could improve by including a parameter that measures distance from top in mm to remove the lesions located in stomach
"""
def hist_compare(data, var):
# helper function to compare corrects vs incorrects for visual comparison
plt.hist([data[data['correct_pred'] == 0][var],
data[data['correct_pred'] == 1][var]],
label=['incorrect', 'correct'],)
def main(args=None):
parser = argparse.ArgumentParser(description='Clustering on NLST Inference By pid and study_yr')
parser.add_argument('--annotated_nodules', help='Path to Directory containing annotated clustered nodules', default='../../../../Google Drive/BDRAD/chest_ct_projects/nodules_annotations_yesh_for_fp_reducer_UnMaskedLungs_2020-06-07.csv')
# parser.add_argument('--annotated_nodules', help='Path to Directory containing annotated clustered nodules', default='../../../../Google Drive/BDRAD/chest_ct_projects/nodules_annotated_yesh_for_fpreducer_MaskedLungs_2020-05-30.csv')
parser.add_argument('--model', help='model to use', default='xgb')
parser = parser.parse_args()
seed = 42
test_size = 0.15
thresh = 0.5
save_fp = './fp_reducer_unmasked.joblib'
print('Saving FP Reducer at: {}'.format(save_fp))
df = pd.read_csv(parser.annotated_nodules, header=0)
df = df[df['correct_pred'].isin(['0',0,'1',1])]
pidyr = df['pid'].astype(str) + '_' + df['study_yr'].astype(str)
print('N Patients: {}'.format(len(np.unique(df['pid']))))
print('N CTs: {}'.format(len(np.unique(pidyr))))
print('N nodules: {}'.format(len(df)))
df['correct_pred'] = df['correct_pred'].astype(int)
df['correct_pred'].mean()
# a bit of preprocessing
df['is_seen_axial_and_coronal'] = np.where(df['is_coronal_mean'].isin([0,1]), 0, 1)
df['dist_to_top'] = np.abs(df['axial_world_coord_mean'] - df['top_world_coord'])
df['dist_to_bottom'] = np.abs(df['axial_world_coord_mean'] - df['bottom_world_coord'])
# split into train and test by PID!!!
train_pids, test_pids = train_test_split(np.unique(df['pid']), test_size=test_size, random_state=seed)
df_train = df[df['pid'].isin(train_pids)].copy()
df_test = df[df['pid'].isin(test_pids)].copy()
# some data exploration
#hist_compare(df_train, 'confidence_max')
# select train vars
train_variables = ['is_seen_axial_and_coronal', 'pred_slice_num_count', 'confidence_max',
'confidence_mean','diameter_max', 'dist_to_top', 'dist_to_bottom']
# Train on best params
if parser.model == 'xgb':
# Grid Search, previously attempted parameters in comments
pos_ratio = sum(df_train['correct_pred']==0) / sum(df_train['correct_pred']==1)
grid_params = grid_params = {'n_estimators': [75, 150, 200], # 1, 2, 4, 5, 6, 7,8,9,10,11,12, 14, 100, 1000
'learning_rate': [0.01, 0.05,0.1,0.5], # 0.01, 0.05, 0.1, 0.3, 0.4, 0.5, 0.6, 0.7, 1
'max_depth': [2,4,6], # 1-4, 6, 8, 10
'subsample': [0.6,0.8,1], # 0.8, 0.9, 1
'colsample_bytree': [0.6,0.8,1], # 0.7, 1
'gamma': [0,1], # 0,1,5
'max_delta_step': [0, 0.1, 0.3], # 0.01, 0.1, 0.3
'scale_pos_weight': [1, np.sqrt(pos_ratio), pos_ratio],
}
model = XGBClassifier(random_state=seed)
model_gridsearch = GridSearchCV(model, grid_params,
cv=StratifiedKFold(n_splits=5, random_state=seed, shuffle=True),
scoring = 'f1', verbose=1, n_jobs=-1)
model_gridsearch.fit(X=df_train[train_variables],
y= df_train['correct_pred'])
params = {**model_gridsearch.best_params_}
print('Best GridSearchCV Score: {}'.format(model_gridsearch.best_score_))
print(params)
model = XGBClassifier(**params)
model.fit(df_train[train_variables], df_train['correct_pred'])
model_full = XGBClassifier(**params)
model_full.fit(df[train_variables], df['correct_pred'])
elif parser.model == 'logreg':
grid_params = {'max_iter':[9999],
'class_weight': [None, 'balanced'],
'penalty':['l2', 'l1']}
model = LogisticRegression()
model = GridSearchCV(model,
grid_params,
cv=StratifiedKFold(n_splits=5, random_state=seed, shuffle=True),
scoring = 'roc_auc', verbose=1, n_jobs=-1)
model.fit(X=df_train[train_variables],
y=df_train['correct_pred'])
print('Best GridSearchCV Score: {}'.format(model.best_score_))
print(model.best_params_)
model_full = LogisticRegression(**model.best_params_)
model_full.fit(df[train_variables], df['correct_pred'])
# Test
print(model)
y_probas = model.predict_proba(df_test[train_variables])[:,1]
y_preds = np.where(y_probas >= thresh, 1, 0)
cm = confusion_matrix(df_test['correct_pred'], y_preds)
cr = classification_report(df_test['correct_pred'], y_preds)
accuracy = accuracy_score(df_test['correct_pred'], y_preds)
rocauc = roc_auc_score(df_test['correct_pred'], y_probas)
print('Accuracy: {}'.format(accuracy))
print('ROC AUC: {}'.format(rocauc))
print(cr)
print('Classes: {}'.format(model.classes_))
print('Confusion Matrix: \n{}'.format(cm))
# print('Model Feature Importances: \n{} \n{}'.format(train_variables, model.feature_importances_))
df_test['pred_score'] = y_probas
df_test_incorrects = df_test[df_test['correct_pred'] != y_preds]
# plot hist
hist_compare(df_test, 'pred_score')
# Plot ROC AUC
fpr, tpr, _ = roc_curve(df_test['correct_pred'], y_probas)
plt.figure()
plt.style.use('ggplot')
lw = 2
plt.plot(fpr, tpr, color='darkorange',
lw=lw, label='AUC = {0:.4f})'.format(rocauc, 3))
plt.plot([0, 1], [0, 1], color='navy', lw=lw, linestyle='--')
plt.xlim([0.0, 1.0])
plt.ylim([0.0, 1.05])
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.title('Receiver Operating Characteristic')
plt.legend(loc="lower right", prop={'size': 12})
plt.show()
# save full trained model
print('-'*30)
print('Save model:')
print(model_full)
dump(model_full, save_fp)
if __name__ == '__main__':
time_total0 = time.time()
main()
print('-'*30)
print('Total Time: {} min'.format(round((time.time()-time_total0)/60, 2)))