def process_image(
image_name: str,
retina_path: Path,
img_output_path: Path,
img_transformed_output_path: Path,
):
retina_img = open_colour_image(retina_path / image_name)
contour = find_eye(retina_img)
# Find bounding box.
x, y, w, h = cv2.boundingRect(contour)
# Crop around bounding box.
retina_img = retina_img[y : y + h, x : x + w]
# Pad to square.
retina_img = pad_to_square(retina_img, w, h, [BLACK, BLACK, BLACK])
# Resize to 1280 x 1280.
retina_img = cv2.resize(retina_img, (1280, 1280), interpolation=cv2.INTER_NEAREST)
transformed_retina_img = transform(retina_img)
new_name = change_suffix(image_name, ".png")
write_image(retina_img, img_output_path / new_name)
write_image(transformed_retina_img, img_transformed_output_path / new_name)
def process_image(i):
path = df["path"].iloc[i]
img = open_colour_image(path)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
contour = find_eye(img)
x, y, w, h = cv2.boundingRect(contour)
img = img[y : y + h, x : x + w]
img = pad_to_square(img, w, h, [BLACK, BLACK, BLACK])
img = cv2.resize(img, (512, 512))
scale = 512
img = cv2.addWeighted(
img, 4, cv2.GaussianBlur(img, (0, 0), scale / 30), -4, 128
)
# Remove outer 10% boundary effects
b = np.zeros(img.shape)
b = cv2.circle(
b,
(img.shape[1] // 2, img.shape[0] // 2),
int(scale * 0.9) // 2,
(1, 1, 1),
-1,
8,
0,
)
img = img * b + 128 * (1 - b)
img = np.transpose(img, (2, 0, 1))
grp_images[i, ...] = img[None]
def process_image(
image_name: str,
retina_path: Path,
ex_path: Path,
he_path: Path,
ma_path: Path,
se_path: Path,
nv_path: Path,
irma_path: Path,
ex_label: np.ndarray,
he_label: np.ndarray,
ma_label: np.ndarray,
se_label: np.ndarray,
nv_label: np.ndarray,
irma_label: np.ndarray,
label_output_path: Path,
inst_output_path: Path,
img_transform_output_path: Path,
od_file_path: str,
colour: bool,
):
retina_img = open_colour_image(retina_path / image_name)
contour = find_eye(retina_img)
ex_img = open_binary_mask(ex_path / image_name)
he_img = open_binary_mask(he_path / image_name)
ma_img = open_binary_mask(ma_path / image_name)
se_img = open_binary_mask(se_path / image_name)
nv_img = open_binary_mask(nv_path / image_name)
irma_img = open_binary_mask(irma_path / image_name)
mask = np.ones((1280, 1280), dtype="uint8") * WHITE
inst = np.ones((1280, 1280), dtype="uint8") * WHITE
fill_contours(mask, [contour], GRAY_CLASS["RETINA"])
draw_od(mask, inst, image_name, od_file_path)
overlay_label(mask, ex_img, ex_label)
overlay_label(mask, he_img, he_label)
overlay_label(mask, ma_img, ma_label)
overlay_label(mask, se_img, se_label)
overlay_label(mask, nv_img, nv_label)
overlay_label(mask, irma_img, irma_label)
transformed_retina_img = transform(retina_img)
if colour:
mask = colour_labels_numpy(mask)
write_image(mask, label_output_path / image_name)
write_image(inst, inst_output_path / image_name)
write_image(transformed_retina_img, img_transform_output_path / image_name)
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"name",
type=str,
)
opt = parser.parse_args()
if opt.name == "copypaste":
generated_retina_path = Path("results") / opt.name / "img"
out_path = Path("results") / opt.name / "transformed"
else:
generated_retina_path = Path("results") / opt.name / "test" / "img"
out_path = Path("results") / opt.name / "test" / "transformed"
out_path.mkdir(parents=True, exist_ok=True)
files = list(generated_retina_path.glob("**/*"))
for path in tqdm(files):
retina_img = open_colour_image(str(path))
transformed_retina_img = transform(retina_img)
write_image(transformed_retina_img, out_path / path.name)
def main():
name = "od_transformed"
out_dir = Path("data/eophtha/od")
out_dir.mkdir(parents=True, exist_ok=True)
model = load_binary_segmentation_model(name)
root_path = Path("/vol/bitbucket/js6317/individual-project/data/e_optha")
path_1 = root_path / "e_optha_MA" / "healthy"
path_2 = root_path / "e_optha_MA" / "MA"
path_3 = root_path / "e_optha_EX" / "EX"
path_4 = root_path / "e_optha_EX" / "healthy"
suffixes = [".JPG", ".jpg", ".PNG", ".png", ".JPEG", ".jpeg"]
files_1 = [f for f in path_1.glob("**/*") if f.is_file() and f.suffix in suffixes]
files_2 = [f for f in path_2.glob("**/*") if f.is_file() and f.suffix in suffixes]
files_3 = [f for f in path_3.glob("**/*") if f.is_file() and f.suffix in suffixes]
files_4 = [f for f in path_4.glob("**/*") if f.is_file() and f.suffix in suffixes]
files = files_1 + files_2 + files_3 + files_4
for file in tqdm(files):
img = open_colour_image(file)
contour = find_eye(img)
# Find bounding box.
x, y, w, h = cv2.boundingRect(contour)
# Crop around bounding box.
img = img[y : y + h, x : x + w]
# Pad to square.
img = pad_to_square(img, w, h, [BLACK, BLACK, BLACK])
# Resize to 1280 x 1280.
img = cv2.resize(img, (1280, 1280), interpolation=cv2.INTER_NEAREST)
transformed_image = transform(img)
predicted = predict_od(model, transformed_image)
image_path = Path(*file.parts[-2:])
new_name = "_".join(image_path.parts)
new_name = change_suffix(new_name, ".png")
cv2.imwrite(f"data/eophtha/od/{new_name}", predicted * 255.0)
def process_image(
retina_path: Path,
ex_path: Path,
ma_path: Path,
od_path: Path,
label_output_path: Path,
instance_output_path: Path,
img_output_path: Path,
img_transformed_output_path: Path,
colour: bool,
):
retina_img = open_colour_image(retina_path)
image_path = Path(*retina_path.parts[-2:])
height, width, _ = retina_img.shape
img_size = (height, width)
contour = find_eye(retina_img)
ex_label = np.ones(img_size, dtype="uint8") * GRAY_CLASS["EX"]
ma_label = np.ones(img_size, dtype="uint8") * GRAY_CLASS["MA"]
ex_img = open_binary_mask(ex_path /
change_suffix(str(image_path), f"_EX.png"),
img_size=img_size)
ma_img = open_binary_mask(ma_path /
change_suffix(str(image_path), f".png"),
img_size=img_size)
mask = np.ones(img_size, dtype="uint8") * WHITE
fill_contours(mask, [contour], GRAY_CLASS["RETINA"])
overlay_label(mask, ex_img, ex_label)
overlay_label(mask, ma_img, ma_label)
inst = np.ones(img_size, dtype="uint8") * WHITE
# Find bounding box.
x, y, w, h = cv2.boundingRect(contour)
# Crop around bounding box.
mask = mask[y:y + h, x:x + w]
inst = inst[y:y + h, x:x + w]
retina_img = retina_img[y:y + h, x:x + w]
# Pad to square.
mask = pad_to_square(mask, w, h, WHITE)
inst = pad_to_square(inst, w, h, WHITE)
retina_img = pad_to_square(retina_img, w, h, [BLACK, BLACK, BLACK])
# Resize to 1280 x 1280.
mask = cv2.resize(mask, (1280, 1280), interpolation=cv2.INTER_NEAREST)
inst = cv2.resize(inst, (1280, 1280), interpolation=cv2.INTER_NEAREST)
retina_img = cv2.resize(retina_img, (1280, 1280),
interpolation=cv2.INTER_NEAREST)
transformed_retina_img = transform(retina_img)
new_name = "_".join(image_path.parts)
new_name = change_suffix(new_name, ".png")
od_img = open_binary_mask(od_path / new_name)
od_label = np.ones((1280, 1280), dtype="uint8") * GRAY_CLASS["OD"]
overlay_label(mask, od_img, od_label)
bg_label = np.ones((1280, 1280), dtype="uint8") * GRAY_CLASS["RETINA"]
overlay_label(inst, od_img, bg_label)
if colour:
mask = colour_labels_numpy(mask)
write_image(mask, label_output_path / new_name)
write_image(inst, instance_output_path / new_name)
write_image(retina_img, img_output_path / new_name)
write_image(transformed_retina_img, img_transformed_output_path / new_name)
def process_image(
image_name: str,
retina_path: Path,
ex_path: Path,
he_path: Path,
ma_path: Path,
se_path: Path,
od_path: Path,
ex_label: np.ndarray,
he_label: np.ndarray,
ma_label: np.ndarray,
se_label: np.ndarray,
od_label: np.ndarray,
bg_label: np.ndarray,
label_output_path: Path,
instance_output_path: Path,
img_output_path: Path,
img_transformed_output_path: Path,
colour: bool,
):
retina_img = open_colour_image(retina_path / image_name)
contour = find_eye(retina_img)
ex_img = open_binary_mask(ex_path / change_suffix(image_name, "_EX.tif"))
he_img = open_binary_mask(he_path / change_suffix(image_name, "_HE.tif"))
ma_img = open_binary_mask(ma_path / change_suffix(image_name, "_MA.tif"))
se_img = open_binary_mask(se_path / change_suffix(image_name, "_SE.tif"))
od_img = open_binary_mask(od_path / change_suffix(image_name, "_OD.tif"))
mask = np.ones((2848, 4288), dtype="uint8") * WHITE
fill_contours(mask, [contour], GRAY_CLASS["RETINA"])
overlay_label(mask, od_img, od_label)
overlay_label(mask, ex_img, ex_label)
overlay_label(mask, he_img, he_label)
overlay_label(mask, ma_img, ma_label)
overlay_label(mask, se_img, se_label)
inst = np.ones((2848, 4288), dtype="uint8") * WHITE
overlay_label(inst, od_img, bg_label)
# Find bounding box.
x, y, w, h = cv2.boundingRect(contour)
# Crop around bounding box.
mask = mask[y:y + h, x:x + w]
inst = inst[y:y + h, x:x + w]
retina_img = retina_img[y:y + h, x:x + w]
# Pad to square.
mask = pad_to_square(mask, w, h, WHITE)
inst = pad_to_square(inst, w, h, WHITE)
retina_img = pad_to_square(retina_img, w, h, [BLACK, BLACK, BLACK])
# Resize to 1280 x 1280.
mask = cv2.resize(mask, (1280, 1280), interpolation=cv2.INTER_NEAREST)
inst = cv2.resize(inst, (1280, 1280), interpolation=cv2.INTER_NEAREST)
retina_img = cv2.resize(retina_img, (1280, 1280),
interpolation=cv2.INTER_NEAREST)
transformed_retina_img = transform(retina_img)
new_name = change_suffix(image_name, ".png")
if colour:
mask = colour_labels_numpy(mask)
write_image(mask, label_output_path / new_name)
write_image(inst, instance_output_path / new_name)
write_image(retina_img, img_output_path / new_name)
write_image(transformed_retina_img, img_transformed_output_path / new_name)
def process_image(
image_name: str,
retina_path: Path,
ex_path: Path,
he_path: Path,
ma_path: Path,
se_path: Path,
retina_mask: np.ndarray,
retina_label: np.ndarray,
ex_label: np.ndarray,
he_label: np.ndarray,
ma_label: np.ndarray,
se_label: np.ndarray,
label_output_path: Path,
inst_output_path: Path,
img_output_path: Path,
img_transformed_output_path: Path,
od_file_path: str,
colour: bool,
):
retina_img = open_colour_image(retina_path / image_name)
contour = find_eye(retina_mask)
ex_img = open_binary_mask(ex_path / image_name)
he_img = open_binary_mask(he_path / image_name)
ma_img = open_binary_mask(ma_path / image_name)
se_img = open_binary_mask(se_path / image_name)
img_size = (1152, 1500)
mask = np.ones(img_size, dtype="uint8") * WHITE
inst = np.ones(img_size, dtype="uint8") * WHITE
thresh = int(MAX_VALUE * CONFIDENCE_THRESHOLD)
overlay_label(mask, retina_mask, retina_label, thresh)
draw_od(mask, inst, image_name, od_file_path)
overlay_label(mask, ex_img, ex_label, thresh=thresh)
overlay_label(mask, he_img, he_label, thresh=thresh)
overlay_label(mask, ma_img, ma_label, thresh=thresh)
overlay_label(mask, se_img, se_label, thresh=thresh)
# Find bounding box.
x, y, w, h = cv2.boundingRect(contour)
# Crop around bounding box.
mask = mask[y : y + h, x : x + w]
inst = inst[y : y + h, x : x + w]
retina_img = retina_img[y : y + h, x : x + w]
# Pad to square.
mask = pad_to_square(mask, w, h, WHITE)
inst = pad_to_square(inst, w, h, WHITE)
retina_img = pad_to_square(retina_img, w, h, [BLACK, BLACK, BLACK])
# Resize to 1280 x 1280.
mask = cv2.resize(mask, (1280, 1280), interpolation=cv2.INTER_NEAREST)
inst = cv2.resize(inst, (1280, 1280), interpolation=cv2.INTER_NEAREST)
retina_img = cv2.resize(retina_img, (1280, 1280), interpolation=cv2.INTER_NEAREST)
transformed_retina_img = transform(retina_img)
if colour:
mask = colour_labels_numpy(mask)
write_image(mask, label_output_path / image_name)
write_image(inst, inst_output_path / image_name)
write_image(retina_img, img_output_path / image_name)
write_image(transformed_retina_img, img_transformed_output_path / image_name)