File size: 6,436 Bytes
a19a7aa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
from torch.utils.data import Dataset
import numpy as np
import random


class FakePatientDataset(Dataset):
    """
    Controllable synthetic multimodal + survival dataset

    You can explicitly control:
    - Final AUC (classification)
    - Final C-index (DFS / OS)
    via interpretable hyperparameters.

    Output: 19 items (aligned with run_epoch_verbose)
    """

    def __init__(
        self,
        n_patients=3000,
        n_slices=30,
        img_size=224,
        num_subtypes=2,
        num_tnm=3,
        seed=2131,

        # =========================
        # ---- AUC controllers ----
        # =========================
        tabular_signal_dims=16,        # ↑ dims β†’ ↑ AUC
        tabular_signal_strength=0.40, # ↑ strength β†’ ↑ AUC
        label_flip_rate=0.10,          # ↑ noise β†’ ↓ AUC

        # =========================
        # ---- C-index controllers
        # =========================
        risk_noise=1.0,               # ↑ noise β†’ ↓ C-index
        dfs_time_noise=6.0,
        os_time_noise=7.0,
        event_sharpness=1.3,          # ↑ β†’ HRζ›΄ζ˜Žζ˜Ύ
    ):
        super().__init__()
        random.seed(seed)
        np.random.seed(seed)

        self.n = n_patients
        self.n_slices = n_slices
        self.img_size = img_size
        self.num_subtypes = num_subtypes
        self.num_tnm = num_tnm

        self.tabular_signal_dims = tabular_signal_dims
        self.tabular_signal_strength = tabular_signal_strength
        self.label_flip_rate = label_flip_rate

        self.risk_noise = risk_noise
        self.dfs_time_noise = dfs_time_noise
        self.os_time_noise = os_time_noise
        self.event_sharpness = event_sharpness

        # =========================
        # Treatment cohort
        # =========================
        self.treatment = np.random.choice(
            [0, 1],
            size=self.n,
            p=[2374 / (2374 + 1790), 1790 / (2374 + 1790)]
        ).astype(np.int64)

        # =========================
        # Ground-truth labels
        # =========================
        self.subtype = np.random.randint(0, num_subtypes, size=self.n).astype(np.int64)
        self.tnm = np.random.randint(0, num_tnm, size=self.n).astype(np.int64)

        # =========================
        # Latent biological risk
        # =========================
        base_risk = (
            0.6 * self.subtype +
            0.5 * self.tnm +
            0.4 * self.treatment +
            np.random.normal(0, self.risk_noise, size=self.n)
        )

        # =========================
        # Survival times
        # =========================
        self.dfs_time = np.clip(
            60 - 7.0 * base_risk + np.random.normal(0, self.dfs_time_noise, size=self.n),
            3, 96
        )
        self.os_time = np.clip(
            75 - 8.5 * base_risk + np.random.normal(0, self.os_time_noise, size=self.n),
            6, 120
        )

        # =========================
        # Event indicators (soft)
        # =========================
        p_dfs = 1 / (1 + np.exp(-(base_risk - 0.2) * self.event_sharpness))
        p_os  = 1 / (1 + np.exp(-(base_risk - 0.4) * self.event_sharpness))

        self.dfs_event = (np.random.rand(self.n) < p_dfs).astype(np.float32)
        self.os_event  = (np.random.rand(self.n) < p_os).astype(np.float32)

        # =========================
        # Time-point labels
        # =========================
        self.dfs_1y = (self.dfs_time <= 12).astype(np.float32)
        self.dfs_3y = (self.dfs_time <= 36).astype(np.float32)
        self.dfs_5y = (self.dfs_time <= 60).astype(np.float32)

        self.os_1y  = (self.os_time <= 12).astype(np.float32)
        self.os_3y  = (self.os_time <= 36).astype(np.float32)
        self.os_5y  = (self.os_time <= 60).astype(np.float32)

    def __len__(self):
        return self.n

    def __getitem__(self, idx):
        s = int(self.subtype[idx])
        t = int(self.tnm[idx])
        tr = int(self.treatment[idx])

        # =========================
        # Label noise (controls AUC ceiling)
        # =========================
        if np.random.rand() < self.label_flip_rate:
            s = 1 - s

        # =========================
        # IMAGE: very weak signal
        # =========================
        base_img = np.random.normal(0.5, 0.30, (self.img_size, self.img_size)).astype(np.float32)
        base_img += 0.03 * s + 0.02 * t + 0.02 * tr
        base_img = np.clip(base_img, 0, 1)

        lesion = torch.from_numpy(
            np.repeat(base_img[None, None, ...], self.n_slices, axis=0)
        )
        space = lesion.clone()

        # =========================
        # TABULAR: main discriminative signal
        # =========================
        radiomics = np.random.normal(0, 1.0, 128).astype(np.float32)
        radiomics[:self.tabular_signal_dims] += (
            self.tabular_signal_strength * s +
            0.7 * self.tabular_signal_strength * t +
            np.random.normal(0, 0.8, self.tabular_signal_dims)
        )

        pet = np.random.normal(0, 1.0, 5).astype(np.float32)
        pet[:2] += 0.5 * self.tabular_signal_strength * s + np.random.normal(0, 0.7, 2)

        clinical = np.random.normal(0, 1.0, 6).astype(np.float32)
        clinical[:3] += 0.5 * self.tabular_signal_strength * t + np.random.normal(0, 0.7, 3)

        return (
            f"P{idx:04d}",

            lesion.float(),
            space.float(),

            torch.from_numpy(radiomics),
            torch.from_numpy(pet),
            torch.from_numpy(clinical),

            torch.tensor(s, dtype=torch.long),
            torch.tensor(t, dtype=torch.long),

            torch.tensor(self.dfs_time[idx], dtype=torch.float32),
            torch.tensor(self.dfs_event[idx], dtype=torch.float32),

            torch.tensor(self.os_time[idx], dtype=torch.float32),
            torch.tensor(self.os_event[idx], dtype=torch.float32),

            torch.tensor(self.dfs_1y[idx], dtype=torch.float32),
            torch.tensor(self.dfs_3y[idx], dtype=torch.float32),
            torch.tensor(self.dfs_5y[idx], dtype=torch.float32),

            torch.tensor(self.os_1y[idx], dtype=torch.float32),
            torch.tensor(self.os_3y[idx], dtype=torch.float32),
            torch.tensor(self.os_5y[idx], dtype=torch.float32),

            torch.tensor(tr, dtype=torch.long),
        )