forked from JimmyCai91/LesionHarvester
-
Notifications
You must be signed in to change notification settings - Fork 0
/
evaluation.py
184 lines (142 loc) · 5.56 KB
/
evaluation.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
# --------------------------------------------------------
# Lesion Harvester
# Licensed under The MIT License [see LICENSE for details]
# Written by Jinzheng Cai
# -------------------------------------------------------
import os
import tqdm
import pickle
import numpy as np
from scipy import interpolate
from voc_eval_lib import voc_ap
def IoU(detected_box, groundtruth_box):
'''Compute overlaps or so called IoU between boxes.'''
# measure the intersection area
ixmin = np.maximum(groundtruth_box[:, 0], detected_box[0])
iymin = np.maximum(groundtruth_box[:, 1], detected_box[1])
ixmax = np.minimum(groundtruth_box[:, 2], detected_box[2])
iymax = np.minimum(groundtruth_box[:, 3], detected_box[3])
iwidth = np.maximum(ixmax - ixmin + 1., 0.)
iheight = np.maximum(iymax - iymin + 1., 0.)
inters = iwidth * iheight
# measure the union area
uni = ((detected_box[2] - detected_box[0] + 1.) * (detected_box[3] - detected_box[1] + 1.) +
(groundtruth_box[:, 2] - groundtruth_box[:, 0] + 1.) *
(groundtruth_box[:, 3] - groundtruth_box[:, 1] + 1.) - inters)
overlaps = inters / np.maximum(1e-6, uni)
return overlaps
def P3DIoU(recist_box, tracklet):
'''
P3D IoU Evaluation Metric
input:
recist_box (tuple): (z, [[x_min, y_min, x_max, y_max]]).
tracklet (dictionary): {z1: [x1_min, y1_min, x1_max, y1_max],
z2: [x2_min, y2_min, x2_max, y2_max], ...}.
algorithm:
if z == z_i and z_i in [z1, z2, ...]:
return IoU([x_min, y_min, x_max, y_max], [xi_min, yi_min, xi_max, yi_max])
else:
return 0
'''
boxes = np.array([xy for _, xy in tracklet])
zs = [z for z, _ in tracklet]
z, xy = recist_box
if z in zs:
ovr = IoU(xy[0], boxes)[zs.index(z)]
else:
ovr = 0
return ovr
if __name__ == '__main__':
p3d_iou_thresh = 0.5
############################
# Part1: load annotation
##############################
annotation_dir = './annotation/Revised-Test1071.pkl'
annotation = pickle.load(open(annotation_dir, 'rb'))
volume_recs = {}
for k in annotation.keys():
volume_recs[k.replace('.annot','')] = []
n_recist = 0
for k, v in annotation.items():
volume_id = k.replace('.annot','')
if len(v) == 0:
continue
z_idxes = np.sort([z for z in v.keys()])
for z in z_idxes:
xys = v[z]
for xy in xys:
volume_recs[volume_id] += [(z, np.array(xy[:4]).reshape((1,-1)))]
n_recist += 1
###############################
# Part2: load detection result
##################################
detections = pickle.load(open('./detection/detectedTest1071.pkl', 'rb'))
volume_ids, det_confs, BB, pred_confs, gts = [], [], [], [], []
for volume_id, data in tqdm.tqdm(detections.items()):
if len(data) == 0:
# No detection in current CT volume
continue
for v in data:
confs, tracklet = [], []
for z, xy in v:
confs.append(xy[-1])
# tracklet.append((z, xy[:-1]))
tracklet.append((z, xy))
BB.append(tracklet)
det_confs.append(np.array(confs).max())
volume_ids.append(volume_id)
overlaps = np.array([P3DIoU(recist, BB[-1]) for recist in volume_recs[volume_ids[-1]]])
if (len(overlaps) > 0) and (overlaps.max() >= p3d_iou_thresh):
gts.append(1)
else:
gts.append(0)
det_confs = np.array(det_confs)
gts = np.array(gts)
###########################
# Part3: do evaluation
###########################
scores = det_confs
ord = np.argsort(scores)[::-1]
nImg = len(volume_recs)
hits = {}
for n, v in volume_recs.items():
hits[n] = np.zeros((len(v),), dtype=bool)
nHits = 0
nMissFPS = 0
nMissAP = 0
nPositive = 0
tps = []
fpsFPS = []
fpsAP = []
for i in ord:
overlaps = np.array([P3DIoU(recist, BB[i]) for recist in volume_recs[volume_ids[i]]])
if (len(overlaps) == 0) or (overlaps.max() < p3d_iou_thresh):
nMissFPS += 1
nMissAP += 1
else:
nPositive += 1
for j in range(len(overlaps)):
if (overlaps[j] >= p3d_iou_thresh):
if (not hits[volume_ids[i]][j]):
hits[volume_ids[i]][j] = True
nHits += 1
else:
# Here we follow AP evalutation in Pascal VOC (https://github.com/rbgirshick/py-faster-rcnn/blob/781a917b378dbfdedb45b6a56189a31982da1b43/lib/datasets/voc_eval.py#L187)
nMissAP += 1
tps.append(nHits)
fpsFPS.append(nMissFPS)
fpsAP.append(nMissAP)
npos = n_recist
rec = np.array(tps) / float(npos)
prec = np.array(tps) / np.maximum(np.array(tps)+np.array(fpsAP), np.finfo(np.float64).eps)
ap = voc_ap(rec, prec, use_07_metric=False)
print('Average precision (AP): {:.4f}'.format(ap))
# FROC used in here (https://github.com/rsummers11/CADLab/tree/master/MULAN_universal_lesion_analysis)
# for evaluating lesion detection in DeepLesion.
# Code for evaluation: https://github.com/rsummers11/CADLab/blob/1192f13b1a6fc0beb3407534a9d3ef7b59df6ba0/lesion_detector_3DCE/rcnn/utils/evaluation.py.
nGt = n_recist
sens = np.array(tps, dtype=float) / nGt
fp_per_img = np.array(fpsFPS, dtype=float) / nImg
f = interpolate.interp1d(fp_per_img, sens, fill_value='extrapolate')
res = f(np.array([0.125, 0.25, 0.5, 1, 2, 4, 8, 16]))
print('\nSensitivity @', [0.125, 0.25, 0.5, 1, 2, 4, 8, 16], '\n average FPs per patient/volume:', ['{:.4f}'.format(re) for re in res])