framework-code-initial-commit

Author: Vision-At-SEECS

config.py

@@ -0,0 +1,113 @@
+from tensorflow.keras.layers import Input
+from easydict import EasyDict as edict
+import tensorflow as tf
+
+config = edict()
+
+# original height and width of image
+config.ORIGINAL_HEIGHT = 2400
+config.ORIGINAL_WIDTH = 1935
+
+# height and width to resize image
+config.HEIGHT = 320
+config.WIDTH  = 256
+
+# input cephalogram image to the base network
+config.IMAGE_INPUT = Input(shape=(config.HEIGHT, config.WIDTH, 3), name="cephalogram")
+
+# landmark region proposals (LRPs) input to the landmark detection network
+config.PROPOSALS_INPUT = Input(shape=(None, 4), name="landmark_region_proposals")
+
+# Image resolution (mm/pixel)
+config.IMAGE_RESOLUTION = 0.1
+
+# cephalometric landmarks
+config.ANATOMICAL_LANDMARKS = {
+    "0": "Sella",
+    "1": "Nasion",
+    "2": "Orbitale",
+    "3": "Porion",
+    "4": "A-point",
+    "5": "B-point",
+    "6": "Pogonion",
+    "7": "Menton",
+    "8": "Gnathion",
+    "9": "Gonion",
+    "10": "Lower Incisal Incision",
+    "11": "Upper Incisal Incision",
+    "12": "Upper Lip",
+    "13": "Lower Lip",
+    "14": "Subnasale",
+    "15": "Soft Tissue Pogonion",
+    "16": "Posterior Nasal Spine",
+    "17": "Anterior Nasal Spine",
+    "18": "Articulare",
+}
+
+# number of cephalometric landmarks
+config.NUM_LANDMARKS = 19
+
+
+config.BACKBONE_BLOCKS_INFO = {
+    "vgg16": {
+        "C1": "block1_conv2",
+        "C2": "block2_conv2",
+        "C3": "block3_conv3",
+        "C4": "block4_conv3",
+        "C5": "block5_conv3"
+    },
+    "vgg19": {
+        "C1": "block1_conv2",
+        "C2": "block2_conv2",
+        "C3": "block3_conv4",
+        "C4": "block4_conv4",
+        "C5": "block5_conv4"
+    },
+    "darknet19": {
+        "C1": "block1_conv1",
+        "C2": "block2_conv1",
+        "C3": "block3_conv3",
+        "C4": "block4_conv3",
+        "C5": "block5_conv5",
+        "C6": "block6_conv5",
+    },
+    "darknet53": {
+        "C1": "block1.1_out",
+        "C2": "block2.2_out",
+        "C3": "block3.8_out",
+        "C4": "block4.8_out",
+        "C5": "block5.4_out"
+    },
+    "resnet18": {
+        "C2": "block2.2_out",
+        "C3": "block3.2_out",
+        "C4": "block4.2_out",
+        "C5": "block5.2_out"
+    },
+    "resnet34": {
+        "C2": "block2.3_out",
+        "C3": "block3.4_out",
+        "C4": "block4.6_out",
+        "C5": "block5.3_out"
+    },
+    "resnet50": {
+        "C2": "conv2_block3_out",
+        "C3": "conv3_block4_out",
+        "C4": "conv4_block6_out",
+        "C5": "conv5_block3_out"
+    }
+}
+
+# Region of interest pool size
+config.ROI_POOL_SIZE = (5, 5)
+
+# margin (in pixels) at each side of lateral skull face
+config.BOX_MARGIN = 32
+
+config.TRAIN = edict()
+# number of epochs
+config.TRAIN.EPOCHS = 10
+# optimizer
+config.TRAIN.OPTIMIZER = tf.keras.optimizers.Adam(learning_rate=0.0001)
+
+cfg = config

data/__init__.py

@@ -0,0 +1,20 @@
+from data.isbi_dataset import ISBIDataset
+from data.pku_dataset import PKUDataset
+from data.dataset import Dataset
+from config import cfg
+
+if __name__ == "__main__":
+    import matplotlib.pyplot as plt
+    import os
+
+    datasets_root_path = "../datasets"
+    isbi_dataset_root = os.path.join(datasets_root_path, "ISBI Dataset")
+
+    data = Dataset(name="ISBI", mode="TRAIN", batch_size=4, shuffle=False)
+
+    images, landmarks = data[0]
+
+    index = 0
+    plt.imshow(images[index])
+    plt.scatter(landmarks[index, :, 0], landmarks[index, :, 1], color="green", s=[3] * cfg.NUM_LANDMARKS)
+    plt.show()

data/dataset.py

@@ -0,0 +1,74 @@
+from data import ISBIDataset, PKUDataset
+from preprocessing import Augmentation
+from paths import Paths
+import tensorflow as tf
+from config import cfg
+import numpy as np
+import cv2
+
+
+def resize(image: np.ndarray, landmarks: np.ndarray):
+    image_height, image_width = image.shape[0:2]
+    ratio_height, ratio_width = (image_height / cfg.HEIGHT), (image_width / cfg.WIDTH)
+
+    image = cv2.resize(np.array(image), dsize=(cfg.WIDTH, cfg.HEIGHT), interpolation=cv2.INTER_CUBIC)
+    landmarks = np.vstack([
+        landmarks[:, 0] / ratio_width,
+        landmarks[:, 1] / ratio_height
+    ]).T
+
+    return image, landmarks
+
+
+class Dataset(tf.keras.utils.Sequence):
+
+    def __init__(
+        self,
+        name: str,
+        mode: str,
+        batch_size: int = 1,
+        augmentation: Augmentation = None,
+        shuffle: bool = False,
+    ):
+
+        if name == "isbi":
+            self.dataset = ISBIDataset(Paths.dataset_root_path(name), mode)
+        elif name == "pku":
+            self.dataset = PKUDataset(Paths.dataset_root_path(name), mode)
+        else:
+            raise ValueError("\'{}\' no such dataset exists in your datasets repository.".format(name))
+
+        self.batch_size = batch_size
+        self.shuffle = shuffle
+
+        if self.shuffle:
+            self.dataset.shuffle()
+
+        self.augmentation = augmentation
+
+    def on_epoch_end(self):
+        if self.shuffle:
+            self.dataset.shuffle()
+
+    def __getitem__(self, index: int):
+
+        start_index = index * self.batch_size
+        end_index = min((index + 1) * self.batch_size, len(self.dataset))
+
+        images = []
+        labels = []
+
+        for index in range(start_index, end_index):
+            image, landmarks = self.dataset[index]
+
+            if self.augmentation is not None:
+                image, landmarks = self.augmentation.apply(image, landmarks)
+
+            image, landmarks = resize(image, landmarks)
+            images.append(image)
+            labels.append(landmarks)
+
+        return tf.stack(images), tf.stack(labels)
+
+    def __len__(self):
+        return len(self.dataset) // self.batch_size

data/isbi_dataset.py

@@ -0,0 +1,71 @@
+from config import cfg
+import numpy as np
+import random
+import cv2
+import os
+
+
+class ISBIDataset(object):
+
+    def __init__(
+        self,
+        dataset_folder_path: str,
+        mode: str
+    ) -> None:
+
+        if mode in ["train", "valid", "test"]:
+            self.mode = mode
+        else:
+            raise ValueError("mode could only be train, valid or test")
+
+        self.imgaes_root_path = os.path.join(dataset_folder_path, self.mode, "images")
+
+        self.annotations_root_path = os.path.join(dataset_folder_path, self.mode, "annotations")
+        self.senior_annotations_root = os.path.join(self.annotations_root_path, "senior-orthodontist")
+        self.junior_annotations_root = os.path.join(self.annotations_root_path, "junior-orthodontist")
+
+        self.images_list = list(sorted(os.listdir(self.imgaes_root_path)))
+
+    def __getitem__(self, index: int):
+        image_file_name = self.images_list[index]
+        label_file_name = self.images_list[index].split(".")[0] + "." + "txt"
+
+        image = self.get_image(image_file_name)
+        label = self.get_label(label_file_name)
+
+        return image, label
+
+    def get_image(self, file_name: str):
+        file_path = os.path.join(self.imgaes_root_path, file_name)
+
+        image = cv2.imread(file_path)
+        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
+
+        return image
+
+    def get_label(self, file_name: str) -> np.ndarray:
+        file_path = os.path.join(self.senior_annotations_root, file_name)
+        with open(file_path) as file:
+            senior_annotations = [landmark.rstrip() for landmark in file]
+
+        senior_annotations = [[float(landmark.split(",")[0]), float(landmark.split(",")[1])] for landmark in senior_annotations[:cfg.NUM_LANDMARKS]]
+        senior_annotations = np.array(senior_annotations, dtype=np.float32)
+
+        file_path = os.path.join(self.junior_annotations_root, file_name)
+        with open(file_path) as file:
+            junior_annotations = [landmark.rstrip() for landmark in file]
+
+        junior_annotations = [[float(landmark.split(",")[0]), float(landmark.split(",")[1])] for landmark in junior_annotations[:cfg.NUM_LANDMARKS]]
+        junior_annotations = np.array(junior_annotations, dtype=np.float32)
+
+        landmarks = np.zeros(shape=(cfg.NUM_LANDMARKS, 2), dtype=np.int32)
+        landmarks[:, 0] = np.ceil((0.5) * (junior_annotations[:, 0] + senior_annotations[:, 0]))
+        landmarks[:, 1] = np.ceil((0.5) * (junior_annotations[:, 1] + senior_annotations[:, 1]))
+
+        return np.array(landmarks, dtype=np.float32)
+
+    def shuffle(self):
+        random.shuffle(self.images_list)
+
+    def __len__(self):
+        return len(self.images_list)

data/pku_dataset.py

@@ -0,0 +1,65 @@
+from config import cfg
+import numpy as np
+import random
+import cv2
+import os
+
+
+class PKUDataset(object):
+
+    def __init__(
+        self,
+        dataset_folder_path: str,
+        mode: str = None
+    ):
+        self.images_root_path = os.path.join(dataset_folder_path, "images")
+        self.labels_root_path = os.path.join(dataset_folder_path, "annotations")
+
+        self.doctor1_annotations_root = os.path.join(self.labels_root_path, "orthodontist-1")
+        self.doctor2_annotations_root = os.path.join(self.labels_root_path, "orthodontist-2")
+
+        self.images_list = os.listdir(self.images_root_path)
+
+    def __getitem__(self, index):
+        image_file_name = self.images_list[index]
+        label_file_name = self.images_list[index].split(".")[0] + "." + "txt"
+
+        image = self.get_image(image_file_name)
+        landmarks = self.get_label(label_file_name)
+
+        return image, landmarks
+
+    def get_image(self, file_name: str):
+        file_path = os.path.join(self.images_root_path, file_name)
+
+        image = cv2.imread(file_path)
+        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
+
+        return np.array(image, dtype=np.uint8)
+
+    def get_label(self, file_name: str):
+        file_path = os.path.join(self.doctor1_annotations_root, file_name)
+        with open(file_path) as file:
+            doctor1_annotations = [landmark.rstrip() for landmark in file]
+
+        doctor1_annotations = [[float(landmark.split(",")[0]), float(landmark.split(",")[1])] for landmark in doctor1_annotations[:cfg.NUM_LANDMARKS]]
+        doctor1_annotations = np.array(doctor1_annotations, dtype=np.float32)
+
+        file_path = os.path.join(self.doctor2_annotations_root, file_name)
+        with open(file_path) as file:
+            doctor2_annotations = [landmark.rstrip() for landmark in file]
+
+        doctor2_annotations = [[float(landmark.split(",")[0]), float(landmark.split(",")[1])] for landmark in doctor2_annotations[:cfg.NUM_LANDMARKS]]
+        doctor2_annotations = np.array(doctor2_annotations, dtype=np.float32)
+
+        landmarks = np.zeros(shape=(cfg.NUM_LANDMARKS, 2), dtype=np.int32)
+        landmarks[:, 0] = np.ceil((0.5) * (doctor1_annotations[:, 0] + doctor2_annotations[:, 0]))
+        landmarks[:, 1] = np.ceil((0.5) * (doctor1_annotations[:, 1] + doctor2_annotations[:, 1]))
+
+        return np.array(landmarks, dtype=np.float32)
+
+    def shuffle(self):
+        random.shuffle(self.images_list)
+
+    def __len__(self):
+        return len(self.images_list)

docs/example.txt

@@ -0,0 +1 @@
+from models.backbone import Backbone