-
Notifications
You must be signed in to change notification settings - Fork 1
/
SAM.py
243 lines (189 loc) · 10.3 KB
/
SAM.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
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
import numpy as np
import scanpy as sc
import pandas as pd
import cv2
import json
import torch
import urllib.request
from tqdm import tqdm
import matplotlib.pyplot as plt
from segment_anything import sam_model_registry, SamAutomaticMaskGenerator, SamPredictor
import supervision as sv
import gseapy
from gseapy import barplot, dotplot
from numba import jit
from log import log
from utils import *
class IAMSAM():
def __init__(self, chkp_path, model_type = 'vit_h'):
DEVICE = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
if DEVICE.type == 'cuda':
print(torch.cuda.get_device_name(0))
print('Memory Usage:')
print('Allocated:', round(torch.cuda.memory_allocated(0)/1024**3,1), 'GB')
print('Cached: ', round(torch.cuda.memory_reserved(0)/1024**3,1), 'GB')
# Download checkpoint if it doesn't exist
if not os.path.exists(chkp_path):
url = 'https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth'
print(f"Downloading checkpoint from {url}...")
with tqdm(unit='B', unit_scale=True, leave=True) as t:
urllib.request.urlretrieve(url, chkp_path, reporthook=lambda *x: t.update(1))
print("Checkpoint downloaded.")
sam = sam_model_registry[model_type](checkpoint=chkp_path)
sam.to(device=DEVICE)
self.predictor = SamPredictor(sam_model=sam)
print("\n\nIAMSAM loaded.")
self.sam = sam
self.masks = []
self.boxes = []
def load_data(self, h5ad_dir):
# Load Anndata
self.adata = sc.read_h5ad(h5ad_dir)
self.adata.var_names_make_unique()
self.adata.X = self.adata.layers['counts']
sc.pp.normalize_total(self.adata, target_sum = 1e4)
sc.pp.log1p(self.adata)
library_id = list(self.adata.uns['spatial'].keys())[0]
# Tissue image(Before crop)
self.tsimg_rgb = cv2.convertScaleAbs(self.adata.uns['spatial'][library_id]['images']['hires']*255)
self.tsimg_bgr = cv2.cvtColor(self.tsimg_rgb, cv2.COLOR_RGB2BGR)
tissue_hires_scalef = self.adata.uns['spatial'][library_id]['scalefactors']['tissue_hires_scalef']
self.adata.obs[['imgcol', 'imgrow']] = self.adata.obsm['spatial']
self.adata.obs['imgrow_'] = tissue_hires_scalef * self.adata.obs['imgrow']
self.adata.obs['imgcol_'] = tissue_hires_scalef * self.adata.obs['imgcol']
self.xrange = [np.min(self.adata.obs['imgcol_']), np.max(self.adata.obs['imgcol_'])]
self.yrange = [np.min(self.adata.obs['imgrow_']), np.max(self.adata.obs['imgrow_'])]
pad = 0.5
self.xrange_ = [round(self.xrange[0]-pad), round(self.xrange[1]+pad)]
self.yrange_ = [round(self.yrange[0]-pad), round(self.yrange[1]+pad)]
# Tissue image (After crop)
self.tsimg_rgb_cropped = self.tsimg_rgb[self.yrange_[0]:self.yrange_[1], self.xrange_[0]:self.xrange_[1]]
self.tsimg_bgr_cropped = self.tsimg_bgr[self.yrange_[0]:self.yrange_[1], self.xrange_[0]:self.xrange_[1]]
log("Image Loaded.")
print('X range : {} ~ {} '.format(self.xrange[0], self.xrange[1]))
print('Y range : {} ~ {} '.format(self.yrange[0], self.yrange[1]))
# For Prompt mode
self.predictor.set_image(self.tsimg_rgb_cropped)
self.prompt_flag = False
# Check cell type information
if not any(self.adata.obs.columns.str.startswith('celltype')):
log("No 'celltype_' column found in adata.obs")
log("Cell type proportion analysis will not be working")
#raise ValueError("No 'celltype_' column found in adata.obs")
def get_mask_prompt_mode(self):
input_boxes = torch.tensor(self.boxes, device=self.predictor.device)
transformed_boxes = self.predictor.transform.apply_boxes_torch(input_boxes, self.tsimg_rgb_cropped.shape[:2])
masks, _, _ = self.predictor.predict_torch(
point_coords=None,
point_labels=None,
boxes=transformed_boxes,
multimask_output=False,
)
mask_list = []
for mask in masks:
mask_ = np.zeros((self.tsimg_rgb.shape[0], self.tsimg_rgb.shape[1])) # Mask size should be original size
mask_[self.yrange_[0]:self.yrange_[1], self.xrange_[0]:self.xrange_[1]] = mask[0].cpu().numpy()
mask_list.append(mask_)
self.masks = mask_list
self.masks_backup = mask_list
masks_integrated = np.zeros((self.tsimg_rgb.shape[0], self.tsimg_rgb.shape[1]))
for ii, mm in enumerate(self.masks):
masks_integrated[mm == 1] = ii + 1 # 1-based index
self.integrated_masks = masks_integrated
return self.masks
def get_mask(self,
box = None,
points_per_side=32,
pred_iou_thresh=0.95, #KEY PARAM : MORE VALUE, LESS CLUSTERS
stability_score_thresh=0.92,
crop_n_layers=1,
crop_n_points_downscale_factor=2,
min_mask_region_area=100 # Requires open-cv to run post-processing
):
mask_generator = SamAutomaticMaskGenerator(
model=self.sam,
points_per_side= points_per_side,
pred_iou_thresh=pred_iou_thresh, # KEY PARAM : MORE VALUE, LESS CLUSTERS
stability_score_thresh=stability_score_thresh,
crop_n_layers=crop_n_layers,
crop_n_points_downscale_factor=crop_n_points_downscale_factor,
min_mask_region_area=min_mask_region_area # Requires open-cv to run post-processing
)
sam_result = mask_generator.generate(self.tsimg_rgb_cropped)
masks_cropped = [mask['segmentation'] for mask in sam_result] # cropped size masks
n_total_masks = len(masks_cropped)
# Get original masks from cropped mask
masks = []
for mask_ in masks_cropped:
mask = np.zeros((self.tsimg_rgb.shape[0], self.tsimg_rgb.shape[1]))
mask[self.yrange_[0]:self.yrange_[1], self.xrange_[0]:self.xrange_[1]] = mask_
masks.append(mask) # original size
# Set on_tissue mask
pixels = np.column_stack((self.adata.obs['imgrow_'].values.astype(int),
self.adata.obs['imgcol_'].values.astype(int)))
tissue_mask = np.zeros((self.tsimg_rgb.shape[0], self.tsimg_rgb.shape[1]))
tissue_mask[pixels[:, 0], pixels[:, 1]] = 1
# filtering out sam_result values that are not on the tissue
on_tissue_sam_result = []
for idx, m in enumerate(masks):
prop = np.sum(cv2.bitwise_and(m, tissue_mask)) / self.adata.n_obs
if prop > 0.001:
on_tissue_sam_result.append(sam_result[idx])
on_tissue_masks = [mask['segmentation'] for mask in sorted(on_tissue_sam_result, key=lambda x: x['area'], reverse=True)] # cropped size
# Get original mask of on_tissue_masks
masks = []
for mask_ in on_tissue_masks:
mask = np.zeros((self.tsimg_rgb.shape[0], self.tsimg_rgb.shape[1]))
mask[self.yrange_[0]:self.yrange_[1], self.xrange_[0]:self.xrange_[1]] = mask_
masks.append(mask)
self.masks = masks
self.masks_backup = masks
mask_annotator = sv.MaskAnnotator()
detections = sv.Detections.from_sam(on_tissue_sam_result)
self.annotated_image = mask_annotator.annotate(self.tsimg_bgr_cropped, detections)
print("{} masks detected, after excluding {} masks not on the tissue".format(
len(on_tissue_masks), n_total_masks - len(on_tissue_masks)))
# Save integrated mask
masks_integrated = np.zeros((self.tsimg_rgb.shape[0], self.tsimg_rgb.shape[1]))
for ii , mm in enumerate(self.masks):
masks_integrated[mm == 1] = ii + 1 # 1-based index
self.integrated_masks = masks_integrated
return self.masks
def extract_degs(self, selected1, selected2, padj_cutoff, lfc_cutoff):
# Add selected mask as 'ROI-1' in adata.obs
roi1_mask = np.zeros((self.tsimg_rgb.shape[0], self.tsimg_rgb.shape[1]))
for idx in selected1:
roi1_mask = roi1_mask + self.masks[idx-1]
roi1_mask = np.array(roi1_mask > 0)
coords = np.round(self.adata.obs[['imgcol_', 'imgrow_']]).astype('int').values
roi1 = calculate_mask_values(coords, roi1_mask)
# if ROI-2 exists then add ROI-2, else add others.
if selected2 is not None:
roi2_mask = np.zeros((self.tsimg_rgb.shape[0], self.tsimg_rgb.shape[1]))
for idx in selected2:
roi2_mask = roi2_mask + self.masks[idx-1]
roi2_mask = np.array(roi2_mask > 0)
roi2 = calculate_mask_values(coords, roi2_mask)
else:
roi2 = np.invert(roi1)
self.adata.obs['ROIs'] = ['ROI1' if in1 else 'ROI2' if in2 else '' for in1, in2 in zip(roi1, roi2)]
# Test DEG
adata_roi = self.adata[np.isin(self.adata.obs['ROIs'], ['ROI1', 'ROI2']),:].copy()
sc.tl.rank_genes_groups(adata_roi, 'ROIs', method='wilcoxon', key_added='DEG')
# DEG_result
self.deg_df = sc.get.rank_genes_groups_df(adata_roi, group = 'ROI1', key = 'DEG')
self.deg_df['-log10Padj'] = -np.log10(self.deg_df['pvals_adj'])
self.deg_df['DE'] = 'None'
self.deg_df.loc[(self.deg_df.pvals_adj < float(padj_cutoff)) & (self.deg_df.logfoldchanges > float(lfc_cutoff)), 'DE'] = 'ROI1'
self.deg_df.loc[(self.deg_df.pvals_adj < float(padj_cutoff)) & (self.deg_df.logfoldchanges < -float(lfc_cutoff)), 'DE'] = 'ROI2'
print("Extract DEGs")
return self.deg_df
# Define the function to calculate the mask values
@jit(nopython=True)
def calculate_mask_values(coords, selmask):
result = np.zeros(coords.shape[0], dtype=np.bool_) # Explicitly set dtype to boolean
for i in range(coords.shape[0]):
x = coords[i, 0]
y = coords[i, 1]
result[i] = selmask[y, x]
return result