parcelcortex/model_apply_parcel.py at master · bthyreau/parcelcortex · GitHub

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
import torch
import nibabel
import numpy as np
import os, sys, time
import torch.nn as nn
import torch.nn.functional as F
from numpy.linalg import inv


class ParcelCortexModel(nn.Module):
    def __init__(self):
        super(ParcelCortexModel, self).__init__()
        self.conv0a_0 = l = nn.Conv3d(1, 16, (1,1,3), padding=(0,0,1))
        self.conv0a_1 = l = nn.Conv3d(16, 16, (1,3,1), padding=(0,1,0))
        l = self.conv0a_2 = nn.Conv3d(16, 16, (3,1,1), padding=(1,0,0))
        l = self.bn0 = nn.BatchNorm3d(16, momentum=1, eps=1e-8)
        l.training = False

        self.maxpool1 = nn.MaxPool3d(2)
        l = self.convf1 = nn.Conv3d(16, 32, (3,3,3), padding=1)
        l = self.bn1 = nn.BatchNorm3d(32, momentum=1, eps=1e-8)
        l.training = False

        l = self.convout2r = nn.Conv3d(32, 96, 1, padding=0)
        self.maxpool2 = nn.MaxPool3d(2)
        l = self.convout2 = nn.Conv3d(96, 96, (3,3,3), padding=1)
        l = self.bn2 = nn.BatchNorm3d(96, momentum=1, eps=1e-8)
        l.training = False

        l = self.convout3r = nn.Conv3d(96, 128, 1, padding=0)
        self.maxpool3 = nn.MaxPool3d(2)
        l = self.convout3p = nn.Conv3d(128, 96, (3,3,3), padding=1)
        l = self.convout3 = nn.Conv3d(96, 128, 1, padding=0)
        l = self.bn3 = nn.BatchNorm3d(128, momentum=1, eps=1e-8)
        l.training = False
        self.maxpool3 = nn.MaxPool3d(2)

        l = self.convlx4 = nn.Conv3d(133, 128, (3,3,3), padding=1)
        l = self.convout4r = nn.Conv3d(128, 128, 1, padding=0)
        l = self.bn4 = nn.BatchNorm3d(128, momentum=1, eps=1e-8)
        l.training = False

        l = self.convlx5 = nn.Conv3d(128, 128, (3,3,3), padding=1)
        l = self.convout5r = nn.Conv3d(128, 128, 1, padding=0)
        l = self.bn5 = nn.BatchNorm3d(128, momentum=1, eps=1e-8)
        l.training = False

        l = self.convlx6 = nn.Conv3d(128, 96, (3,3,3), padding=1)
        l = self.convout6r = nn.Conv3d(96+96, 96, 1, padding=0)
        l = self.bn6 = nn.BatchNorm3d(96, momentum=1, eps=1e-8)
        l.training = False

        l = self.convlx7 = nn.Conv3d(96, 96, (3,3,3), padding=1)
        l = self.convout7r = nn.Conv3d(96+32, 96, 1, padding=0)
        l = self.bn7 = nn.BatchNorm3d(96, momentum=1, eps=1e-8)
        l.training = False

        self.conv8a_0 = l = nn.Conv3d(96, 96, (1,1,3), padding=(0,0,1))
        self.conv8a_1 = l = nn.Conv3d(96, 96, (1,3,1), padding=(0,1,0))
        l = self.conv8a_2 = nn.Conv3d(96, 96, (3,1,1), padding=(1,0,0))
        l = self.convlx8 = nn.Conv3d(96+16, 75, 1, padding=0)

    def forward(self, x, atlas_hint, side_hint):
        x = F.relu(self.conv0a_0(x))
        x = F.relu(self.conv0a_1(x))
        x = self.conv0a_2(x)
        x = F.relu(self.bn0(x))
        li0 = x

        x = self.maxpool1(x)
        x = self.convf1(x)
        x = F.relu(self.bn1(x))
        li1 = x

        x = F.relu(self.convout2r(x))
        x = self.maxpool2(x)
        x = self.convout2(x)
        x = F.relu(self.bn2(x))
        li2 = x

        x = F.relu(self.convout3r(x))
        x = self.maxpool3(x)
        x = F.relu(self.convout3p(x))
        x = self.convout3(x)
        x = F.relu(self.bn3(x))
        x = self.maxpool3(x)

        atlas_hint = F.interpolate(atlas_hint[...,None,None,None], (3,6,6), mode="nearest")
        side_hint = F.interpolate(side_hint[...,None,None,None], (3,6,6), mode="nearest")
        x = torch.cat([x, atlas_hint, side_hint], dim=1)

        x = F.relu(self.convlx4(x))
        x = (self.convout4r(x))
        x = F.relu(self.bn4(x))

        x = F.interpolate(x, scale_factor=2, mode="nearest")
        x = F.relu(self.convlx5(x))
        x = (self.convout5r(x))
        x = F.relu(self.bn5(x))

        x = F.interpolate(x, scale_factor=2, mode="nearest")
        x = F.relu(self.convlx6(x))
        x = torch.cat([x, li2], dim=1)
        x = (self.convout6r(x))
        x = F.relu(self.bn6(x))


        x = F.interpolate(x, scale_factor=2, mode="nearest")
        x = F.relu(self.convlx7(x))
        x = torch.cat([x, li1], dim=1)
        x = (self.convout7r(x))
        x = F.relu(self.bn7(x))


        x = F.interpolate(x, scale_factor=2, mode="nearest")
        x = F.relu(self.conv8a_0(x))
        x = F.relu(self.conv8a_1(x))
        x = F.relu(self.conv8a_2(x))

        x = torch.cat([x, li0], dim=1)
        x = self.convlx8(x)
        x = torch.sigmoid(x)
        return x

net = ParcelCortexModel()
net.eval()
# os.path.dirname(os.path.realpath(__file__)) + "/parcelcortex.pt")

net.load_state_dict(torch.load(os.path.dirname(os.path.realpath(__file__)) + "/parcelcortex.pt"))


atlas_codes = {"a2009": ([1,0,0], 75), "aseg": ([0, 1, 0], 35), "pals": ([0, 0, 1], 48)}
hemi_template_file = os.path.dirname(os.path.realpath(__file__)) + "/templates/dil_ig_ribbon_ig_b96_box128_lout_T1_thr.nii.gz"
roi = nibabel.load(hemi_template_file).get_fdata() > .5

if len(sys.argv) > 1:
    fnamel = sys.argv[1]
    assert("b96_box128_lout" in fnamel)

    if len(sys.argv) >= 3:
        atlas_list = sys.argv[2:]
        assert all([atlas in ["a2009", "aseg", "pals"] for atlas in atlas_list])
    else:
        atlas_list = ["a2009", "aseg", "pals"]

    T = time.time()
    for atlas in atlas_list:
        print("Applying model for atlas %s" % atlas)
        for fname in [fnamel, fnamel.replace("_lout_", "_rout_")]:
            img = nibabel.load(fname)

            d = img.get_fdata(dtype=np.float32)
            if d.max() > 10:
                d /= 255. # d is probably uchar encoded
            d_orr = d
            side_hint = [1, 0]

            if "_rout_" in fname:
                d_orr = d_orr[::-1].copy() # copy because pytorch fail at negative strides
                side_hint = side_hint[::-1]

            #print("Starting inference on %s using atlas %s" % (fname, atlas))
            atlas_code, nb_roi = atlas_codes[atlas]
            d_orr[~roi] = 0
            with torch.no_grad():
                out1 = net(torch.Tensor(d_orr[None,None]), torch.Tensor([atlas_code]), torch.Tensor([side_hint]))
            #print("Inference " + str(time.time() - T))
            out1 = np.asarray(out1)
            a=np.argmax(out1[:,:nb_roi], axis=1) + 1
            a[out1[:,:nb_roi].max(axis=1) < .001] = 0 # mostly for debug
            outt = a[0].astype(np.uint8)
            outt[~roi] = 0 # no need to fill too far

            if "_rout_" in fname:
                outt = outt[::-1]

            nibabel.Nifti1Image(outt, img.affine).to_filename( fname.replace(".nii.gz", "_outlab_%s_filled.nii.gz" % atlas))