Rydhi commited on
Commit
67e6a2e
·
1 Parent(s): 8a5f1cf

Pick slice with most tumor instead of fixed middle slice

Browse files
Files changed (1) hide show
  1. inference.py +40 -19
inference.py CHANGED
@@ -3,35 +3,40 @@ import numpy as np
3
  import torch
4
  import matplotlib.pyplot as plt
5
  from matplotlib.colors import ListedColormap
6
- from monai.inferers import sliding_window_inference
7
  from monai.data import MetaTensor
8
  from monai.transforms import Orientation, EnsureType
 
9
  from sklearn.preprocessing import MinMaxScaler
10
 
11
-
12
-
13
  SW_ROI = [128, 128, 128]
14
  SW_BATCH_SIZE = 2
15
 
16
- def load_nifti(file_path):
17
- img = nib.load(file_path)
18
- return img.get_fdata().astype(np.float32), img.affine
19
 
 
 
 
20
 
21
  def normalize(x: np.ndarray) -> np.ndarray:
22
  scaler = MinMaxScaler(feature_range=(0, 1))
23
  normalized_1D_array = scaler.fit_transform(x.reshape(-1, x.shape[-1]))
24
  return normalized_1D_array.reshape(x.shape)
25
 
 
26
  def reorient_to_ras(x: MetaTensor) -> MetaTensor:
27
  return Orientation(axcodes="RAS")(x)
28
 
 
29
  def convert_metatensor_to_numpy(x: MetaTensor) -> np.ndarray:
30
  return EnsureType(data_type="numpy", track_meta=False)(x)
31
 
 
32
  def crop_brats2021_volume(x: np.ndarray) -> np.ndarray:
33
  return x[:, 56:184, 56:184, 13:141]
34
 
 
35
  def preprocess_single_modality(fp):
36
  nifti_data = nib.load(fp)
37
  data = nifti_data.get_fdata()
@@ -46,8 +51,9 @@ def preprocess_single_modality(fp):
46
  data = crop_brats2021_volume(data) # -> (1, 128, 128, 128)
47
  return data
48
 
 
49
  def preprocess_volume(t1_path, t1ce_path, t2_path, flair_path):
50
- # match training channel order: Flair, T1w, T1gd, T2w
51
  flair = preprocess_single_modality(flair_path).swapaxes(1, 3)
52
  t1w = preprocess_single_modality(t1_path).swapaxes(1, 3)
53
  t1gd = preprocess_single_modality(t1ce_path).swapaxes(1, 3)
@@ -56,8 +62,13 @@ def preprocess_volume(t1_path, t1ce_path, t2_path, flair_path):
56
  volume = np.concatenate((flair, t1w, t1gd, t2w), axis=0).astype(np.float32)
57
  return volume, None
58
 
 
 
 
 
 
59
  def run_inference(model, volume, device):
60
- tensor = torch.from_numpy(volume).unsqueeze(0).to(device)
61
 
62
  with torch.no_grad():
63
  output = sliding_window_inference(
@@ -69,23 +80,35 @@ def run_inference(model, volume, device):
69
  )
70
  pred = (torch.sigmoid(output) > 0.5).float()
71
 
72
- return pred[0].cpu().numpy()
73
 
74
- region_cmap = ListedColormap(["none", "yellow", "green", "red"])
 
 
 
75
 
76
  def combine_regions(vol_3ch):
77
  tc, wt, et = vol_3ch[0], vol_3ch[1], vol_3ch[2]
78
  combined = np.zeros_like(wt)
79
- combined[wt > 0] = 1
80
- combined[tc > 0] = 2
81
- combined[et > 0] = 3
82
  return combined
83
 
 
 
 
 
 
 
 
 
 
 
84
  def make_slice_figure(volume, pred, save_path="preview.png"):
85
- mid = volume.shape[1] // 2
86
- flair_slice = volume[0][mid] # channel 0 = FLAIR (matches your training order)
87
 
88
- pred_combined = combine_regions(pred)[mid]
89
  pred_tc, pred_wt, pred_et = pred[0][mid], pred[1][mid], pred[2][mid]
90
 
91
  rows = [
@@ -112,9 +135,8 @@ def make_slice_figure(volume, pred, save_path="preview.png"):
112
  return save_path
113
 
114
 
115
-
116
  def make_combined_figure(volume, pred, save_path="combined.png"):
117
- mid = volume.shape[1] // 2
118
  flair_slice = volume[0][mid]
119
  pred_combined = combine_regions(pred)[mid]
120
 
@@ -124,7 +146,6 @@ def make_combined_figure(volume, pred, save_path="combined.png"):
124
  ax.set_title("Combined Prediction", fontsize=14)
125
  ax.axis("off")
126
 
127
- from matplotlib.patches import Patch
128
  legend_elements = [
129
  Patch(facecolor="yellow", label="Whole Tumor (WT)"),
130
  Patch(facecolor="green", label="Tumor Core (TC)"),
 
3
  import torch
4
  import matplotlib.pyplot as plt
5
  from matplotlib.colors import ListedColormap
6
+ from matplotlib.patches import Patch
7
  from monai.data import MetaTensor
8
  from monai.transforms import Orientation, EnsureType
9
+ from monai.inferers import sliding_window_inference
10
  from sklearn.preprocessing import MinMaxScaler
11
 
 
 
12
  SW_ROI = [128, 128, 128]
13
  SW_BATCH_SIZE = 2
14
 
15
+ region_cmap = ListedColormap(["none", "yellow", "green", "red"])
16
+
 
17
 
18
+ # ---------------------------------------------------------------------------
19
+ # Preprocessing (matches training pipeline exactly)
20
+ # ---------------------------------------------------------------------------
21
 
22
  def normalize(x: np.ndarray) -> np.ndarray:
23
  scaler = MinMaxScaler(feature_range=(0, 1))
24
  normalized_1D_array = scaler.fit_transform(x.reshape(-1, x.shape[-1]))
25
  return normalized_1D_array.reshape(x.shape)
26
 
27
+
28
  def reorient_to_ras(x: MetaTensor) -> MetaTensor:
29
  return Orientation(axcodes="RAS")(x)
30
 
31
+
32
  def convert_metatensor_to_numpy(x: MetaTensor) -> np.ndarray:
33
  return EnsureType(data_type="numpy", track_meta=False)(x)
34
 
35
+
36
  def crop_brats2021_volume(x: np.ndarray) -> np.ndarray:
37
  return x[:, 56:184, 56:184, 13:141]
38
 
39
+
40
  def preprocess_single_modality(fp):
41
  nifti_data = nib.load(fp)
42
  data = nifti_data.get_fdata()
 
51
  data = crop_brats2021_volume(data) # -> (1, 128, 128, 128)
52
  return data
53
 
54
+
55
  def preprocess_volume(t1_path, t1ce_path, t2_path, flair_path):
56
+ # channel order matches training: Flair, T1w, T1gd, T2w
57
  flair = preprocess_single_modality(flair_path).swapaxes(1, 3)
58
  t1w = preprocess_single_modality(t1_path).swapaxes(1, 3)
59
  t1gd = preprocess_single_modality(t1ce_path).swapaxes(1, 3)
 
62
  volume = np.concatenate((flair, t1w, t1gd, t2w), axis=0).astype(np.float32)
63
  return volume, None
64
 
65
+
66
+ # ---------------------------------------------------------------------------
67
+ # Inference
68
+ # ---------------------------------------------------------------------------
69
+
70
  def run_inference(model, volume, device):
71
+ tensor = torch.from_numpy(volume).unsqueeze(0).to(device) # (1, 4, D, H, W)
72
 
73
  with torch.no_grad():
74
  output = sliding_window_inference(
 
80
  )
81
  pred = (torch.sigmoid(output) > 0.5).float()
82
 
83
+ return pred[0].cpu().numpy() # (3, D, H, W) -- channel order: TC, WT, ET
84
 
85
+
86
+ # ---------------------------------------------------------------------------
87
+ # Visualization helpers
88
+ # ---------------------------------------------------------------------------
89
 
90
  def combine_regions(vol_3ch):
91
  tc, wt, et = vol_3ch[0], vol_3ch[1], vol_3ch[2]
92
  combined = np.zeros_like(wt)
93
+ combined[wt > 0] = 1 # whole tumor / edema region -> yellow
94
+ combined[tc > 0] = 2 # tumor core -> green
95
+ combined[et > 0] = 3 # enhancing tumor -> red
96
  return combined
97
 
98
+
99
+ def find_best_slice(pred):
100
+ """Pick the D-axis slice with the most total tumor voxels (TC+WT+ET),
101
+ so small regions like TC/ET are actually visible instead of relying
102
+ on the raw geometric middle slice."""
103
+ combined_mask = (pred[0] > 0) | (pred[1] > 0) | (pred[2] > 0)
104
+ voxels_per_slice = combined_mask.sum(axis=(1, 2))
105
+ return int(voxels_per_slice.argmax())
106
+
107
+
108
  def make_slice_figure(volume, pred, save_path="preview.png"):
109
+ mid = find_best_slice(pred)
110
+ flair_slice = volume[0][mid] # channel 0 = FLAIR
111
 
 
112
  pred_tc, pred_wt, pred_et = pred[0][mid], pred[1][mid], pred[2][mid]
113
 
114
  rows = [
 
135
  return save_path
136
 
137
 
 
138
  def make_combined_figure(volume, pred, save_path="combined.png"):
139
+ mid = find_best_slice(pred)
140
  flair_slice = volume[0][mid]
141
  pred_combined = combine_regions(pred)[mid]
142
 
 
146
  ax.set_title("Combined Prediction", fontsize=14)
147
  ax.axis("off")
148
 
 
149
  legend_elements = [
150
  Patch(facecolor="yellow", label="Whole Tumor (WT)"),
151
  Patch(facecolor="green", label="Tumor Core (TC)"),