Spaces:
Sleeping
Sleeping
Commit
·
2ef318a
1
Parent(s):
273da2c
Upload app.py
Browse files
app.py
ADDED
@@ -0,0 +1,151 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import cv2
|
2 |
+
|
3 |
+
import numpy as np
|
4 |
+
import streamlit as st
|
5 |
+
import tensorflow as tf
|
6 |
+
|
7 |
+
from utils import _get_retina_bb, _pad_to_square
|
8 |
+
|
9 |
+
|
10 |
+
@st.cache(allow_output_mutation=True)
|
11 |
+
def load_model(model_file):
|
12 |
+
model = tf.keras.models.load_model(model_file, compile=False)
|
13 |
+
print(f'Model {model_file} Loaded!')
|
14 |
+
return model
|
15 |
+
|
16 |
+
|
17 |
+
@st.cache(allow_output_mutation=True)
|
18 |
+
def load_gatekeeper():
|
19 |
+
validator_model = tf.keras.models.load_model('checkpoints/ResNetV2-EyeQ-QA.tf')
|
20 |
+
print('Gatekeeper Model Loaded!')
|
21 |
+
return validator_model
|
22 |
+
|
23 |
+
|
24 |
+
def parse_function(image):
|
25 |
+
image = tf.image.resize(image, [512, 512])
|
26 |
+
image = tf.image.convert_image_dtype(image, tf.float32)
|
27 |
+
return image
|
28 |
+
|
29 |
+
|
30 |
+
def main():
|
31 |
+
|
32 |
+
st.title('Retina Segmentation')
|
33 |
+
|
34 |
+
st.sidebar.title('Segmentation Model')
|
35 |
+
options = st.sidebar.selectbox('Select Option:', ('Vessels', 'Lesions (BETA)'))
|
36 |
+
gatekeeper = st.sidebar.radio("Gatekeeper:", ('Enabled', 'Disabled'))
|
37 |
+
|
38 |
+
gatekeeper_model = load_gatekeeper()
|
39 |
+
|
40 |
+
if options == 'Vessels':
|
41 |
+
|
42 |
+
st.set_option('deprecation.showfileUploaderEncoding', False)
|
43 |
+
uploaded_file = st.file_uploader('Choose an image...', type=('png', 'jpg', 'jpeg'))
|
44 |
+
|
45 |
+
model = load_model('checkpoints/DeeplabV3Plus_DRIVE.tf')
|
46 |
+
|
47 |
+
if uploaded_file:
|
48 |
+
col1, col2 = st.columns(2)
|
49 |
+
|
50 |
+
# Load Image
|
51 |
+
file_bytes = np.asarray(bytearray(uploaded_file.read()), dtype=np.uint8)
|
52 |
+
image = cv2.imdecode(file_bytes, 1)
|
53 |
+
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
|
54 |
+
|
55 |
+
# Check image
|
56 |
+
valid = np.argmax(gatekeeper_model(parse_function(image[None, ...])))
|
57 |
+
if valid == 2 and gatekeeper == 'Enabled':
|
58 |
+
st.image(image)
|
59 |
+
st.info('Image is of poor quality')
|
60 |
+
return
|
61 |
+
|
62 |
+
# Localise and center retina image
|
63 |
+
x, y, w, h, _ = _get_retina_bb(image)
|
64 |
+
image = image[y:y + h, x:x + w, :]
|
65 |
+
image = _pad_to_square(image, border=0)
|
66 |
+
image = cv2.resize(image, (1024, 1024))
|
67 |
+
|
68 |
+
with col1:
|
69 |
+
st.subheader("Uploaded Image")
|
70 |
+
st.image(image)
|
71 |
+
|
72 |
+
# Apply CLAHE pre-processing
|
73 |
+
clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(16, 16))
|
74 |
+
image = cv2.cvtColor(image, cv2.COLOR_RGB2LAB)
|
75 |
+
image[:, :, 0] = clahe.apply(image[:, :, 0])
|
76 |
+
image = cv2.cvtColor(image, cv2.COLOR_LAB2RGB)
|
77 |
+
image = cv2.cvtColor(image, cv2.COLOR_RGB2GRAY)
|
78 |
+
image = tf.image.convert_image_dtype(image, tf.float32)
|
79 |
+
|
80 |
+
# Run model on input
|
81 |
+
y_pred = model(image[None, ..., None])[0].numpy()
|
82 |
+
|
83 |
+
with col2:
|
84 |
+
st.subheader("Predicted Vessel")
|
85 |
+
st.image(y_pred)
|
86 |
+
|
87 |
+
elif options == 'Lesions (BETA)':
|
88 |
+
|
89 |
+
st.write('```--- WARNING: This model is highly experimental ---```')
|
90 |
+
|
91 |
+
st.set_option('deprecation.showfileUploaderEncoding', False)
|
92 |
+
uploaded_file = st.file_uploader('Choose an image...', type=('png', 'jpg', 'jpeg'))
|
93 |
+
|
94 |
+
model = load_model('checkpoints/DeeplabV3Plus_FGADR.tf')
|
95 |
+
|
96 |
+
if uploaded_file:
|
97 |
+
col1, col2, col3, = st.columns(3)
|
98 |
+
|
99 |
+
# Load Image
|
100 |
+
file_bytes = np.asarray(bytearray(uploaded_file.read()), dtype=np.uint8)
|
101 |
+
image = cv2.imdecode(file_bytes, 1)
|
102 |
+
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
|
103 |
+
|
104 |
+
# Check image
|
105 |
+
valid = np.argmax(gatekeeper_model(parse_function(image[None, ...])))
|
106 |
+
if valid == 2 and gatekeeper == 'Enabled':
|
107 |
+
st.image(image)
|
108 |
+
st.info('Image is of poor quality')
|
109 |
+
return
|
110 |
+
|
111 |
+
# Localise and center retina image
|
112 |
+
x, y, w, h, _ = _get_retina_bb(image)
|
113 |
+
image = image[y:y + h, x:x + w, :]
|
114 |
+
image = _pad_to_square(image, border=0)
|
115 |
+
image = cv2.resize(image, (1024, 1024))
|
116 |
+
|
117 |
+
with col1:
|
118 |
+
st.subheader("Uploaded Image")
|
119 |
+
st.image(image)
|
120 |
+
|
121 |
+
# Apply CLAHE pre-processing
|
122 |
+
clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(16, 16))
|
123 |
+
image = cv2.cvtColor(image, cv2.COLOR_RGB2LAB)
|
124 |
+
image[:, :, 0] = clahe.apply(image[:, :, 0])
|
125 |
+
image = cv2.cvtColor(image, cv2.COLOR_LAB2RGB)
|
126 |
+
image = tf.image.convert_image_dtype(image, tf.float32)
|
127 |
+
|
128 |
+
# Run model on input
|
129 |
+
y_pred = model(image[None, ..., None])[0].numpy()
|
130 |
+
|
131 |
+
with col2:
|
132 |
+
st.subheader(f'MA')
|
133 |
+
st.image(y_pred[..., 1])
|
134 |
+
with col3:
|
135 |
+
st.subheader(f'HE')
|
136 |
+
st.image(y_pred[..., 2])
|
137 |
+
with col1:
|
138 |
+
st.subheader(f'EX')
|
139 |
+
st.image(y_pred[..., 3])
|
140 |
+
with col2:
|
141 |
+
st.subheader(f'SE')
|
142 |
+
st.image(y_pred[..., 4])
|
143 |
+
with col3:
|
144 |
+
st.subheader(f'OD')
|
145 |
+
st.image(y_pred[..., 5])
|
146 |
+
|
147 |
+
if __name__ == '__main__':
|
148 |
+
|
149 |
+
tf.config.set_visible_devices([], 'GPU')
|
150 |
+
|
151 |
+
main()
|
utils.py
ADDED
@@ -0,0 +1,137 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import cv2
|
2 |
+
|
3 |
+
import matplotlib.pyplot as plt
|
4 |
+
import numpy as np
|
5 |
+
|
6 |
+
|
7 |
+
def _pad_to_square(image, long_side=None, border=0):
|
8 |
+
h, w, _ = image.shape
|
9 |
+
|
10 |
+
if long_side == None: long_side = max(h, w)
|
11 |
+
|
12 |
+
l_pad = (long_side - w) // 2 + border
|
13 |
+
r_pad = (long_side - w) // 2 + border
|
14 |
+
t_pad = (long_side - h) // 2 + border
|
15 |
+
b_pad = (long_side - h) // 2 + border
|
16 |
+
if w % 2 != 0: r_pad = r_pad + 1
|
17 |
+
if h % 2 != 0: b_pad = b_pad + 1
|
18 |
+
|
19 |
+
image = np.pad(
|
20 |
+
image,
|
21 |
+
((t_pad, b_pad),
|
22 |
+
(l_pad, r_pad),
|
23 |
+
(0, 0)),
|
24 |
+
'constant')
|
25 |
+
|
26 |
+
return image
|
27 |
+
|
28 |
+
|
29 |
+
def _get_retina_bb(image):
|
30 |
+
|
31 |
+
# make image greyscale and normalise
|
32 |
+
image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
|
33 |
+
image = cv2.normalize(image, None, 0, 255, cv2.NORM_MINMAX)
|
34 |
+
|
35 |
+
# calculate threshold perform threshold
|
36 |
+
threshold = np.mean(image)/3-7
|
37 |
+
ret, thresh = cv2.threshold(image, max(0, threshold), 255, cv2.THRESH_BINARY)
|
38 |
+
|
39 |
+
# median filter, erode and dilate to remove noise and holes
|
40 |
+
thresh = cv2.medianBlur(thresh, 25)
|
41 |
+
thresh = cv2.erode(thresh, None, iterations=2)
|
42 |
+
thresh = cv2.dilate(thresh, None, iterations=2)
|
43 |
+
|
44 |
+
# find mask contour
|
45 |
+
cnts = cv2.findContours(thresh, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
|
46 |
+
cnts = cnts[0] if len(cnts) == 2 else cnts[1]
|
47 |
+
c = max(cnts, key=cv2.contourArea)
|
48 |
+
|
49 |
+
# Get bounding box from mask contour
|
50 |
+
x, y, w, h = cv2.boundingRect(c)
|
51 |
+
|
52 |
+
# Get mask from contour
|
53 |
+
mask = np.zeros_like(image)
|
54 |
+
cv2.drawContours(mask, [c], -1, (255, 255, 255), -1)
|
55 |
+
|
56 |
+
return x, y, w, h, mask
|
57 |
+
|
58 |
+
|
59 |
+
def _get_retina_bb2(image, skips=4):
|
60 |
+
'''
|
61 |
+
Experimental Retina Bounding Box detector based on Convexity Defect Points
|
62 |
+
'''
|
63 |
+
|
64 |
+
# make image greyscale and normalise
|
65 |
+
image = cv2.cvtColor(image, cv2.COLOR_RGB2GRAY)
|
66 |
+
image = cv2.normalize(image, None, 0, 255, cv2.NORM_MINMAX)
|
67 |
+
|
68 |
+
# calculate threshold perform threshold
|
69 |
+
threshold = np.mean(image)/3-7
|
70 |
+
ret, thresh = cv2.threshold(image, max(0, threshold), 255, cv2.THRESH_BINARY)
|
71 |
+
|
72 |
+
# median filter, erode and dilate to remove noise and holes
|
73 |
+
thresh = cv2.medianBlur(thresh, 25)
|
74 |
+
thresh = cv2.erode(thresh, None, iterations=2)
|
75 |
+
thresh = cv2.dilate(thresh, None, iterations=2)
|
76 |
+
|
77 |
+
# Find contours
|
78 |
+
contours, hierarchy = cv2.findContours(thresh, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)
|
79 |
+
|
80 |
+
# Find the index of the largest contour
|
81 |
+
areas = [cv2.contourArea(c) for c in contours]
|
82 |
+
max_index = np.argmax(areas)
|
83 |
+
cnt = contours[max_index]
|
84 |
+
|
85 |
+
# Get convexity defect points
|
86 |
+
hull = cv2.convexHull(cnt, returnPoints=False)
|
87 |
+
hull[::-1].sort(axis=0)
|
88 |
+
defects = cv2.convexityDefects(cnt, hull)
|
89 |
+
|
90 |
+
ConvexityDefectPoint = []
|
91 |
+
for i in range(0, defects.shape[0], skips):
|
92 |
+
s, e, f, d = defects[i, 0]
|
93 |
+
ConvexityDefectPoint.append(tuple(cnt[f][0]))
|
94 |
+
|
95 |
+
# Get minimum enclosing circle as retina estimate
|
96 |
+
(x, y), radius = cv2.minEnclosingCircle(np.array(ConvexityDefectPoint))
|
97 |
+
|
98 |
+
# Get mask from contour
|
99 |
+
mask = np.zeros_like(image)
|
100 |
+
cv2.circle(mask, (x, y), radius, (255, 255, 255), -1)
|
101 |
+
|
102 |
+
# return (x, y, w, h) bounding box
|
103 |
+
return int(x - radius), int(y - radius), int(2 * radius - 1), int(2 * radius - 1), mask
|
104 |
+
|
105 |
+
|
106 |
+
def rgb_clahe(image, clipLimit=2.0, tileGridSize=(16, 16)):
|
107 |
+
lab = cv2.cvtColor(image, cv2.COLOR_RGB2LAB)
|
108 |
+
clahe = cv2.createCLAHE(clipLimit=clipLimit, tileGridSize=tileGridSize)
|
109 |
+
lab[..., 0] = clahe.apply(lab[..., 0])
|
110 |
+
return cv2.cvtColor(lab, cv2.COLOR_LAB2RGB)
|
111 |
+
|
112 |
+
|
113 |
+
if __name__ == '__main__':
|
114 |
+
|
115 |
+
# image_file = '/vol/biomedic3/bh1511/retina/IDRID/segmentation/0_Original_Images/IDRiD_65.jpg'
|
116 |
+
# image_file = '/vol/biomedic3/bh1511/retina/CHASE_DB1/images/Image_08R.jpg'
|
117 |
+
# image_file = '/vol/vipdata/data/retina/kaggle-diabetic-retinopathy-detection/train/16_right.jpeg'
|
118 |
+
# image_file = '/vol/vipdata/data/retina/IDRID/a_segmentation/images/train/IDRiD_01.jpg'
|
119 |
+
image_file = 'preprocessing/Image_10L.png'
|
120 |
+
|
121 |
+
# Load Image
|
122 |
+
image = cv2.imread(image_file)
|
123 |
+
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
|
124 |
+
plt.imshow(image); plt.show()
|
125 |
+
|
126 |
+
# Localise and center retina image
|
127 |
+
x, y, w, h, _ = _get_retina_bb(image)
|
128 |
+
image = image[y:y + h, x:x + w, :]
|
129 |
+
image = _pad_to_square(image, border=0)
|
130 |
+
image = cv2.resize(image, (1024, 1024))
|
131 |
+
|
132 |
+
# Apply CLAHE pre-processing
|
133 |
+
image = rgb_clahe(image)
|
134 |
+
|
135 |
+
# Display or save image
|
136 |
+
plt.imshow(image); plt.show()
|
137 |
+
# cv2.imwrite('processed_image.png', image)
|