def custom_crop_last(slide_, y, batch_to_predict, valid_bit_matrix, x_max):
batch_to_predict[y] = []
valid_bit_matrix[y] = []
for x in range(0, x_max - 1):
image_crop = slide.read_slide_crop(
slide_, x * CROP_SIZE,
slide.get_slide_size(slide_)[1] * CROP_SIZE).convert('RGB')
resized_crop = slide.resize_image_a(image_crop, 224, 224)
resized_crop_np = np.asarray(resized_crop, np.float32)
if filter.check_valid(resized_crop_np):
valid_bit_matrix[y].append(1)
else:
valid_bit_matrix[y].append(0)
batch_to_predict[y].append(resized_crop_np)
image_crop = slide.read_slide_crop(
slide_,
slide.get_slide_size(slide_)[0] - CROP_SIZE,
slide.get_slide_size(slide_)[1] * CROP_SIZE).convert('RGB')
resized_crop = slide.resize_image_a(image_crop, 224, 224)
resized_crop_np = np.asarray(resized_crop, np.float32)
if filter.check_valid(resized_crop_np):
valid_bit_matrix[y].append(1)
else:
valid_bit_matrix[y].append(0)
batch_to_predict[y].append(resized_crop_np)
def balance_dataset(x, y, p):
cropped_dataset_folder = path.join(CROP_FOLDER, str(1120))
unique, counts = np.unique(y, return_counts=True)
max_class_size = max(counts)
for i in range(0, unique.shape[0]):
if counts[i] < max_class_size:
fileList = glob.glob(
path.join(cropped_dataset_folder,
"*" + CATEGORIES[unique[i]] + "*"))
for j in range(0, max_class_size - counts[i]):
if len(fileList) > 0:
filename = random.choice(fileList)
fileList.remove(filename)
img_patient = os.path.basename(filename).split("_")[0]
img_class = unique[i]
img = Image.open(filename).convert('RGB')
img_array = np.asarray(img, np.uint8)
if filter.check_valid(img_array):
x.append(img_array)
y.append(img_class)
p.append(img_patient)
else:
log.print_warning("Img " + filename +
" not loaded: too much white")
j = j - 1
else:
log.print_warning("No more available images for class " +
CATEGORIES[unique[i]])
break
return x, y, p
def load_datasets(*sampSizes):
x = []
y = []
p = []
for ss in sampSizes:
log.print_debug("Opening Cropped dataset " + str(ss))
cropped_dataset_folder = path.join(CROP_FOLDER, str(ss))
for filename in os.listdir(cropped_dataset_folder):
try:
img_path = path.join(cropped_dataset_folder, filename)
img_patient = filename.split("_")[0]
img_class = CATEGORIES.index(str(filename.split("_")[1]))
img = Image.open(img_path).convert('RGB')
img_array = np.asarray(img, np.uint8)
if filter.check_valid(img_array):
to_append_img = np.asarray(
img.resize((int(INPUT_SIZE), int(INPUT_SIZE)),
Image.LANCZOS))
x.append(to_append_img)
y.append(img_class)
p.append(img_patient)
else:
log.print_warning("Img " + filename +
" not loaded: too much white")
except Exception as e:
log.print_error("Cannot load image " + filename)
return x, y, p
def custom_crop(slide_, y, batch_to_predict, valid_bit_list, x_max, crop_size):
for x in range(0, x_max - 1):
image_crop = slide.read_slide_crop(slide_, x * crop_size,
y * crop_size,
crop_size).convert('RGB')
resized_crop = slide.resize_image_a(image_crop, 224, 224)
resized_crop_np = np.asarray(resized_crop, np.float32)
if filter.check_valid(resized_crop_np):
valid_bit_list[y * x_max + x] = 1
else:
valid_bit_list[y * x_max + x] = 0
batch_to_predict[y * x_max + x] = resized_crop_np
image_crop = slide.read_slide_crop(
slide_,
slide.get_slide_size(slide_)[0] - crop_size, y * crop_size,
crop_size).convert('RGB')
resized_crop = slide.resize_image_a(image_crop, 224, 224)
resized_crop_np = np.asarray(resized_crop, np.float32)
if filter.check_valid(resized_crop_np):
valid_bit_list[y * x_max + x_max - 1] = 1
else:
valid_bit_list[y * x_max + x_max - 1] = 0
batch_to_predict[y * x_max + x_max - 1] = resized_crop_np
def balance_set(x, y, in_set_patients):
log.print_debug("Balancing dataset")
cropped_dataset_folder = path.join(CROP_FOLDER, str(1120))
x_list = x.tolist()
unique, counts = np.unique(y, return_counts=True)
max_class_size = max(counts)
for i in range(0, unique.shape[0]):
if counts[i] < max_class_size:
file_list = glob.glob(
path.join(cropped_dataset_folder,
"*" + CATEGORIES[unique[i]] + "*"))
cleaned_file_list = file_list.copy()
for filename in cleaned_file_list:
img_patient = os.path.basename(filename).split("_")[0]
if img_patient in in_set_patients:
cleaned_file_list.remove(filename)
images_to_add = max_class_size - counts[i]
for j in range(0, max_class_size - counts[i]):
if len(cleaned_file_list) > 0:
filename = random.choice(cleaned_file_list)
cleaned_file_list.remove(filename)
img_class = unique[i]
img = Image.open(filename).convert('RGB')
img_array = np.asarray(img, np.uint8)
if filter.check_valid(img_array):
to_append_img = np.asarray(
img.resize((int(INPUT_SIZE), int(INPUT_SIZE)),
Image.LANCZOS))
x_list.append(to_append_img)
y = np.append(y, img_class)
images_to_add = images_to_add - 1
#log.print_debug("Img " + filename + " added to set. " + str( images_to_add ) + " images to go.")
else:
log.print_warning("Img " + filename +
" not loaded: too much white")
continue
else:
log.print_warning("No more available images for class " +
CATEGORIES[unique[i]])
break
return np.asarray(x_list), y