Datasets:

ArXiv:
File size: 10,402 Bytes
36fdbcf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import nibabel as nib
import numpy as np
import torch
import torch.nn.functional as F
import surface_distance
from surface_distance import metrics
from src.utils.util import save_predict, save_csv



def model_predict(args, img, prompt, img_encoder, prompt_encoder, mask_decoder):
    patch_size = args.rand_crop_size[0]
    device = args.device
    out = F.interpolate(img.float(), scale_factor=512 / patch_size, mode='trilinear')
    input_batch = out[0].transpose(0, 1)
    batch_features, feature_list = img_encoder(input_batch)
    feature_list.append(batch_features)
    #feature_list = feature_list[::-1]
    points_torch = prompt.transpose(0, 1)
    new_feature = []
    for i, (feature, feature_decoder) in enumerate(zip(feature_list, prompt_encoder)):
        if i == 3:
            new_feature.append(
                feature_decoder(feature.to(device), points_torch.clone(), [patch_size, patch_size, patch_size])
            )
        else:
            new_feature.append(feature.to(device))
    img_resize = F.interpolate(img[0, 0].permute(1, 2, 0).unsqueeze(0).unsqueeze(0).to(device), scale_factor=64/patch_size,
                               mode="trilinear")
    new_feature.append(img_resize)
    masks = mask_decoder(new_feature, 2, patch_size//64)
    masks = masks.permute(0, 1, 4, 2, 3)
    return masks

def get_points_prompt(args, points_dict, cumulative=False):
    """
    get prompt tensor (input) with given point-locations
    """
    patch_size = args.rand_crop_size[0]

    # the first point is always same, and we use it as anchor point, i.e. (x[0], y[0], z[0]) --> (206, 126, 320)
    # manually test with 1 prompt, 5 prompts and 10 prompts
    x, y, z = points_dict['x_location'], points_dict['y_location'], points_dict['z_location'] # x with size--> tensor([num_prompts, 1])

    x_m = (torch.max(x) + torch.min(x)) // 2
    y_m = (torch.max(y) + torch.min(y)) // 2
    z_m = (torch.max(z) + torch.min(z)) // 2

    # considered transpose in dataloader, e.g. match to real original images
    d_min = x_m - patch_size // 2
    d_max = x_m + patch_size // 2
    h_min = z_m - patch_size // 2
    h_max = z_m + patch_size // 2
    w_min = y_m - patch_size // 2
    w_max = y_m + patch_size // 2



    points = torch.cat([z - d_min, x - w_min, y - h_min], dim=1).unsqueeze(1).float()
    points_torch = points.to(args.device)

    patch_dict = {'w_min': w_min, 'w_max': w_max, 'h_min': h_min, 'h_max': h_max, 'd_min': d_min, 'd_max': d_max}

    return points_torch, patch_dict



def get_final_prediction(args, img, seg_dict, points_dict, img_encoder, prompt_encoder_list, mask_decoder):
    seg = seg_dict['seg']

    device = args.device
    patch_size = args.rand_crop_size[0]

    points_torch, patch_dict = get_points_prompt(args, points_dict)


    w_min, w_max = patch_dict['w_min'], patch_dict['w_max']
    h_min, h_max = patch_dict['h_min'], patch_dict['h_max']
    d_min, d_max = patch_dict['d_min'], patch_dict['d_max']



    w_l = max(0, -w_min)
    w_r = max(0, w_max - points_dict['z_dimension'])
    h_l = max(0, -h_min)
    h_r = max(0, h_max - points_dict['y_dimension'])
    d_l = max(0, -d_min)
    d_r = max(0, d_max - points_dict['x_dimension'])

    d_min = max(0, d_min)
    h_min = max(0, h_min)
    w_min = max(0, w_min)

    img_patch = img[:, :, d_min:d_max, h_min:h_max, w_min:w_max].clone()
    prompt_patch = seg_dict['prompt'][:, :, d_min:d_max, h_min:h_max, w_min:w_max].clone()
    print(torch.unique(prompt_patch))
    print(prompt_patch.sum())

    l = len(torch.where(prompt_patch == 1)[0][10:])
    sample = np.random.choice(np.arange(l), l, replace=True)
    x_negative, y_negative, z_negative = get_points(seg_dict['prompt'], sample, positive=True)
    points_negative = torch.cat([z_negative - d_min, x_negative - w_min, y_negative - h_min], dim=1).unsqueeze(1).float().to(args.device)
    points_torch = torch.cat([points_torch, points_negative], dim=0)
    #points_torch = points_torch[:1]
    print(points_torch.size())
    # the pad follows the format (pad ordering) of torch
    # pad deals the points that are not included in the original patch (default size: 128^3)
    # if such case happens, the points values are negative
    # global query in prompt encoder is a learnable variable
    # this may because the point values represent the coordinate information with a shifted origin (-d_min, -h_min, -w_min)
    img_patch = F.pad(img_patch, (w_l, w_r, h_l, h_r, d_l, d_r)) # follow the format of torch pad function

    pred = model_predict(args,
                         img_patch,
                         points_torch,
                         img_encoder,
                         prompt_encoder_list,
                         mask_decoder)
    pred = pred[:, :, d_l:patch_size - d_r, h_l:patch_size - h_r, w_l:patch_size - w_r]
    pred = F.softmax(pred, dim=1)[:, 1]
    # seg_pred is not meta tensor
    # seg_pred = torch.zeros(1, 1, points_dict['x_dimension'], points_dict['z_dimension'], points_dict['y_dimension']).to(device)
    # zeros_like carries meta tensor
    seg_pred = torch.zeros_like(img).to(device)[:, 0, :].unsqueeze(0)
    seg_pred[:, :, d_min:d_max, h_min:h_max, w_min:w_max] += pred

    final_pred = F.interpolate(seg_pred, size=seg.shape[2:], mode="trilinear")
    img_orig = F.interpolate(img, size=seg.shape[2:], mode="trilinear")

    # original_image = F.pad(input=img, pad=(0, 1, 0, 0), mode='constant', value=0)
    return final_pred, img_orig


def get_points(prompt, sample, positive=True):
    value = 1 if positive else 0
    z = torch.where(prompt == value)[3][sample].unsqueeze(1)  # --> tensor([[x_value]]) instead of tensor([x_value])
    x = torch.where(prompt == value)[2][sample].unsqueeze(1)  # ignore: check datasets.py, --> self.spatial_index
    y = torch.where(prompt == value)[4][sample].unsqueeze(1)
    # consider x,y,z here is a,b,c. and abc need to match features after img_encoder with size (b,c,x,z,y)
    # e.g. a-->x, b-->z, c-->y
    # xyz are the coordinates after dataloader, e.g. without consider the self.spatial_index
    return x, y, z
def get_points_location(args, prompt, negative=False):
    """
    use this to get anchor points
    """
    np.random.seed(args.seed)
    l = len(torch.where(prompt == 1)[0])
    sample = np.random.choice(np.arange(l), args.num_prompts, replace=True)
    x, y, z = get_points(prompt, sample)
    if negative:
        l = len(torch.where(prompt == 0)[0])
        sample = np.random.choice(np.arange(l), args.num_prompts, replace=True)
        x_negative, y_negative, z_negative = get_points(prompt, sample, positive=False)
        x = torch.cat([x, x_negative], dim=0)
        y = torch.cat([y, y_negative], dim=0)
        z = torch.cat([z, z_negative], dim=0)

    # this x, y, z location follows the original after change spatial_index
    points_dict = {'x_location': x, 'y_location': y, 'z_location': z,
                   'x_dimension': prompt.shape[2], 'y_dimension': prompt.shape[3], 'z_dimension': prompt.shape[4]}
    return points_dict
def get_input(args, img, seg, negative=False):
    seg = seg.float().unsqueeze(0)
    seg = seg.to(args.device)
    img = img.to(args.device)
    prompt = F.interpolate(seg, img.shape[2:], mode="nearest")
    points_dict = get_points_location(args, prompt, negative=negative)
    seg_dict = {'seg': seg, 'prompt': prompt}
    return img, seg_dict, points_dict


def calculate_cost(args,
                   prediction, ground_truth,
                   loss_function, spacing,
                   loss_list, loss_nsd_list):

    loss = 1 - loss_function(prediction, ground_truth)
    loss_value = loss.squeeze(0).squeeze(0).squeeze(0).squeeze(0).squeeze(0).detach().cpu().numpy()

    ssd = surface_distance.compute_surface_distances(
        (ground_truth == 1)[0, 0].cpu().numpy(),
        (prediction == 1)[0, 0].cpu().numpy(),
        spacing_mm=spacing[0].numpy()
    )
    nsd = metrics.compute_surface_dice_at_tolerance(ssd, args.tolerance)

    loss_list.append(loss_value)
    loss_nsd_list.append(nsd)

    return loss_list, loss_nsd_list, loss_value, nsd


def tester(args, logger,
        model_dict, test_data, loss_list, loss_nsd_list,
        loss_function):
    img_encoder, prompt_encoder_list, mask_decoder = model_dict['img_encoder'], model_dict['prompt_encoder_list'], \
    model_dict['mask_decoder']

    patient_list = []

    for idx, (img, seg, spacing) in enumerate(test_data):
        print( 'current / total subjects:  {} / {}'
               .format(idx + 1, len(test_data.dataset.img_dict)))
        image_path = test_data.dataset.img_dict[idx]
        image_data = nib.load(image_path)

        if args.data == 'pancreas':
            patient_name = test_data.dataset.img_dict[idx].split('/')[-2] + '.nii.gz'
        else:
            patient_name = test_data.dataset.img_dict[idx].split('/')[-1]

        patient_list.append(patient_name)

        img, seg_dict, points_dict = get_input(args, img, seg, negative=False) # get input and point prompt

        # pred
        final_pred, img_orig_space = get_final_prediction(args, img, seg_dict, points_dict, img_encoder,
                                                                      prompt_encoder_list, mask_decoder)

        masks = final_pred > 0.5
        loss_list, loss_nsd_list, loss, nsd = calculate_cost(args,
                                                  masks, seg_dict['seg'],
                                                  loss_function, spacing,
                                                  loss_list, loss_nsd_list)
        # print(1 - loss_function(masks, seg))
        logger.info(
            " Case {}/{} {} - Dice {:.6f} | NSD {:.6f}".format(
                idx + 1, len(test_data.dataset.img_dict), test_data.dataset.img_dict[idx], loss.item(), nsd))

        if args.save_predictions:
            save_predict(args, logger, final_pred, seg_dict['seg'], masks, points_dict, idx, test_data, image_data, patient_name) # coordinates are recorded in the log file


    # all subject
    mean_dice, mean_nsd = np.mean(loss_list), np.mean(loss_nsd_list)
    logger.info("- Test metrics Dice: " + str(mean_dice))
    logger.info("- Test metrics NSD: " + str(mean_nsd))
    logger.info("----------------------")

    if args.save_csv:
        save_csv(args, logger, patient_list, loss_list, loss_nsd_list)

    return loss_list, loss_nsd_list, patient_list