{"cells":[{"metadata":{},"cell_type":"markdown","source":"\n
\n

VinBigData Chest X-ray Abnormalities Detection

\n

Automatically localize and classify thoracic abnormalities from chest radiographs

\n
\n
\n
\n
\n

About Competition

\n

\n Radiologists diagnose and treat medical conditions using imaging techniques like CT and PET scans, MRIs, and, of course, X-rays. Yet, as it happens when working with such a wide variety of medical tools, radiologists face many daily challenges, perhaps the most difficult being the chest radiograph. The interpretation of chest X-rays can lead to medical misdiagnosis, even for the best practicing doctor. Computer-aided detection and diagnosis systems (CADe/CADx) would help reduce the pressure on doctors at metropolitan hospitals and improve diagnostic quality in rural areas.\n

\n

\n In this competition we are to predict the thoracic abnormalities in given X-Ray images and also locate those abnormalities. The data provided include:\n

\n

\n
\n
\n\n
\n

Table of Contents

\n
\n First Look at the Data1\n EDA2\n An insight of the Data2\n Data Preparation3\n Model Building and training4\n
\n
\n
\n

Note:

\n
The utilities_x_ray module used here is a script that I have written(can be found here). It contains some functions for visualization of the X-Ray images. The dicom image reading pipeline is taken from this Notebook by @raddar
"},{"metadata":{"_uuid":"8f2839f25d086af736a60e9eeb907d3b93b6e0e5","_cell_guid":"b1076dfc-b9ad-4769-8c92-a6c4dae69d19","trusted":true},"cell_type":"code","source":"import numpy as np\nimport random\nimport tensorflow as tf\nimport tensorflow.keras.layers as L\nimport tensorflow.keras.backend as K\nimport seaborn as sns\nimport pandas as pd\nimport cv2\nimport os\nimport matplotlib.pyplot as plt\nfrom utilities_x_ray import read_xray,showXray\nfrom tqdm import tqdm\nimport pydicom\n\nimport warnings\nwarnings.filterwarnings(\"ignore\")\n\nfrom keras import preprocessing, layers\nfrom keras.models import Sequential\n\nfrom sklearn.preprocessing import MultiLabelBinarizer\nfrom sklearn.model_selection import train_test_split\nimport os\nimport cv2\nfrom PIL import Image\nimport pickle\nfrom pathlib import Path\n\nimport imageio\nimport pydicom\nfrom pydicom.pixel_data_handlers.util import apply_voi_lut\n\nimport os","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"# Import Libraries\nimport matplotlib.pyplot as plt\nimport torch\nfrom torch import nn\nfrom torch import optim\nimport torch.nn.functional as F\nfrom torchvision import datasets, transforms, models\nfrom torch.autograd import Variable","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"### To test whether GPU instance is present in the system of not.\nuse_cuda = torch.cuda.is_available()\nprint('Using PyTorch version:', torch.__version__, 'CUDA:', use_cuda)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"

First Look at the data

     "},{"metadata":{},"cell_type":"markdown","source":"## 1. DataFrames"},{"metadata":{"_uuid":"d629ff2d2480ee46fbb7e2d37f6b5fab8052498a","_cell_guid":"79c7e3d0-c299-4dcb-8224-4455121ee9b0","trusted":true,"_kg_hide-input":true},"cell_type":"code","source":"df_train = pd.read_csv('../input/vinbigdata-chest-xray-abnormalities-detection/train.csv')\nss = pd.read_csv('../input/vinbigdata-chest-xray-abnormalities-detection/sample_submission.csv')","execution_count":null,"outputs":[]},{"metadata":{"_kg_hide-input":true,"trusted":true},"cell_type":"code","source":"df_train.head()","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"ids = df_train.image_id[:10]\nimgs = []\nlabels = []\nnps = []\nim_label_set = []\nn_images = 100\n\ntrain = os.listdir('../input/vinbigdata-chest-xray-abnormalities-detection/train')\ntest = os.listdir('../input/vinbigdata-chest-xray-abnormalities-detection/test')\n\ntest_img = ('../input/vinbigdata-chest-xray-abnormalities-detection/test/')\ntrain_img = ('../input/vinbigdata-chest-xray-abnormalities-detection/train/')","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":""},{"metadata":{"trusted":true},"cell_type":"code","source":"ss.head()","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"The submission file must contain the image id and the prediction string in the format \"a b (c,d,e,f)\"
where\n"},{"metadata":{},"cell_type":"markdown","source":"## 2. Images"},{"metadata":{"trusted":true},"cell_type":"code","source":"plt.figure(figsize=(8,10))\nplt.imshow(read_xray('../input/vinbigdata-chest-xray-abnormalities-detection/train/0108949daa13dc94634a7d650a05c0bb.dicom'),cmap=plt.cm.bone)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"

EDA

     "},{"metadata":{"_kg_hide-input":true,"trusted":true},"cell_type":"code","source":"print(\"Number of rows in train dataframe: {}\".format(df_train.shape[0]))\nprint(\"Number of Unique images in train set: {}\".format(df_train.image_id.nunique()))\nprint(\"Number of Classes: {}\\n\".format(df_train.class_name.nunique()))\nprint(\"Class Names: {}\".format(list(df_train.class_name.unique())))","execution_count":null,"outputs":[]},{"metadata":{"_kg_hide-input":true,"trusted":true},"cell_type":"code","source":"print(\"Null Values:\")\ndf_train.isna().sum().to_frame().rename(columns={0:'Null Value count'}).style.background_gradient('viridis')","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"The number of null values are same as the number of samples that do not have any abnormality"},{"metadata":{},"cell_type":"markdown","source":"### The Distribution of Classes\nWe can see there is a huge class imbalance. The number of negative examples are very high and a few abnormalities have very few examples "},{"metadata":{"_kg_hide-input":true,"trusted":true},"cell_type":"code","source":"plt.figure(figsize=(9,6))\nsns.countplot(df_train[\"class_id\"]);\nplt.title(\"Class Distributions\");","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"### Distribution of Radiologists"},{"metadata":{"_kg_hide-input":true,"trusted":true},"cell_type":"code","source":"plt.figure(figsize=(9,6))\nsns.countplot(df_train[\"rad_id\"]);\nplt.title(\"rad_id Distributions\");","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"

An Intuition of the Data

     

\n
Before proceeding further let us try and get an intuition of the data and what exactly we need to do.
\n
In this competition we have been given 15000 images for training. Parallelly we have a dataframe containing the ground truths for various abnormalities. Every sample in the datframe contains:
\n \nImportant points to be noted here are:\n\n

Information about dicom can be found: Here

\n

Procedure to extract DICOM metadata can be found in: this notebook

"},{"metadata":{},"cell_type":"markdown","source":"

Data Preparation

     "},{"metadata":{"trusted":true},"cell_type":"code","source":"class_names = sorted(df_train.class_name.unique())\ndel class_names[class_names.index('No finding')]\nclass_names = class_names+['No finding']\nclasses = dict(zip(list(range(15)),class_names))","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"def prepareDataFrame(train_df= df_train):\n train_df = train_df.fillna(0)\n cols = ['image_id','label']+list(range(4*len(class_names[:-1])))\n return_df = pd.DataFrame(columns=cols)\n \n for image in tqdm(train_df.image_id.unique()):\n df = train_df.query(\"image_id==@image\")\n label = np.zeros(15)\n for cls in df.class_id.unique():\n label[int(cls)]=1\n bboxes_df = df.groupby('class_id')[['x_min','y_min','x_max','y_max']].mean().round()\n \n bboxes_list = [0 for i in range(60)]\n for ind in list(bboxes_df.index):\n bboxes_list[4*ind:4*ind+4] = list(bboxes_df.loc[ind,:].values)\n return_df.loc[len(return_df),:] = [image]+[label]+bboxes_list[:-4]\n return return_df\ntrain_df = prepareDataFrame()","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"train_df.head(2)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"temp=train_df[train_df.image_id=='50a418190bc3fb1ef1633bf9678929b3']\nnp.array([temp.iloc[0,1],temp.iloc[0,2:].values])","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"### To test whether GPU instance is present in the system of not.\nuse_cuda = torch.cuda.is_available()\nprint('Using PyTorch version:', torch.__version__, 'CUDA:', use_cuda) ","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"device = torch.device(\"cuda\" if use_cuda else \"cpu\")\ndevice","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"class DataLoader:\n def __init__(self,path = None,train_df=train_df):\n self.path = path\n self.files = os.listdir(self.path)\n np.random.shuffle(self.files)\n self.df = train_df\n \n def read_image(self):\n for img in self.files:\n im_name = img.split('.dicom')[0]\n image = read_xray(self.path+img)\n image = cv2.resize(image,(256,256),cv2.INTER_AREA)\n image = np.expand_dims(image,axis=2)\n temp = self.df[self.df.image_id==im_name]\n c_label,bb = temp.iloc[0,1],temp.iloc[0,2:].values.astype('float')\n yield image,c_label,bb\n \n \n def batch_generator(self,items,batch_size):\n a=[]\n i=0\n for item in items:\n a.append(item)\n i+=1\n\n if i%batch_size==0:\n yield a\n a=[]\n if len(a) is not 0:\n yield a\n \n def flow(self,batch_size):\n \"\"\"\n flow from given directory in batches\n ==========================================\n batch_size: size of the batch\n \"\"\"\n while True:\n for bat in self.batch_generator(self.read_image(),batch_size):\n batch_images = []\n batch_c_labels = []\n batch_bb = []\n for im,im_c_label,im_bb in bat:\n batch_images.append(im)\n batch_c_labels.append(im_c_label)\n batch_bb.append(im_bb)\n batch_images = np.stack(batch_images,axis=0)\n batch_labels = (np.stack(batch_c_labels,axis=0),np.stack(batch_bb,axis=0))\n yield batch_images,batch_labels","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"dl = DataLoader('../input/vinbigdata-chest-xray-abnormalities-detection/train/')\ntrain_set = dl.flow(batch_size=32)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"

Model Building and Training

     "},{"metadata":{"trusted":true},"cell_type":"code","source":"for n, id_ in enumerate(ids):\n dicom_path = train_img + id_ + '.dicom'\n dicom = pydicom.read_file(dicom_path)\n data = apply_voi_lut(dicom.pixel_array, dicom)\n \n data = data - np.min(data)\n \n data = data / np.max(data)\n \n data = (data * 255).astype(np.uint8)\n new_shape = tuple([int(x / 3) for x in data.shape])\n data = cv2.resize(data, (new_shape[1], new_shape[0]))\n im = Image.fromarray(data)\n \n #imgs.append(im)\n \n new_im = im.resize((256,256))\n npdata = np.asarray(new_im)\n new = preprocessing.image.img_to_array(npdata)\n nps.append(new)\n \n #im.save('test'+str(n)+'.jpg')\n \n df = df_train[df_train.image_id == id_]\n label = df.iloc[0,2]\n #label = tf.io.decode_raw(label, tf.uint8)\n #label = tf.reshape(label, label.shape)\n #label = tf.one_hot(label, 10)\n labels.append(label) ","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"# compile data into a keras readable format\n\nlabels = np.stack(labels)\ndata = np.stack(nps)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"dl = DataLoader('../input/vinbigdata-chest-xray-abnormalities-detection/train/')\ntrain_set = dl.flow(batch_size=32)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"def build_cnn():\n model = tf.keras.models.Sequential()\n model.add(tf.keras.layers.Conv2D(32, (3, 3), activation='relu',strides=(2,2), input_shape=(256, 256, 3)))\n model.add(tf.keras.layers.MaxPooling2D((2, 2)))\n model.add(tf.keras.layers.Conv2D(64, (3, 3), strides=(2,2),activation='relu'))\n model.add(tf.keras.layers.MaxPooling2D((2, 2)))\n model.add(tf.keras.layers.Conv2D(128, (3, 3), strides=(2,2),activation='relu'))\n model.add(tf.keras.layers.Flatten())\n model.add(tf.keras.layers.Dense(128, activation='relu'))\n model.add(tf.keras.layers.Dense(15, activation='softmax'))\n return model","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"# Build CNN model\nmodel = build_cnn()\n#Compile the model with optimizer and loss function\nmodel.compile(optimizer='adam',loss='categorical_crossentropy',\nmetrics=['accuracy'])","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"# split data into train & test\n\nX_train, X_test, y_train, y_test = train_test_split(data, labels, test_size=0.2, random_state=34)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"batch_size = 256\nepochs = 50\nhistory = model.fit(X_train, y_train,\n batch_size=batch_size,\n epochs=epochs,\n verbose=0)\nscore = model.evaluate(X_test, y_test, verbose=0)\n","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"# create ohe of labels\n\ncomp_labels = [0,1,2,3,4,5,6,7,8,9,10,11,12,13,14]\n\nmlb = MultiLabelBinarizer()\n\nmlb.fit([labels])","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"# verify all classes are present\n\nmlb.classes_","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"# examine X_train shape before feeding model \n\nX_train.shape","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"# view model summary to get an idea of its parameters and steps\n\nmodel.summary()","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"# Save the CNN model to disk for later use.\nmodel_path = \"models/pneumiacnn\"\nmodel.save(filepath=model_path)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"# evaluate model on the given input data from only 3000 of the training images\n\nmodel.evaluate(X_test, y_test)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"# visualize model steps \n\ntf.keras.utils.plot_model(model)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"test_df = pd.DataFrame(results, columns=['image_id', 'PredictionString'])\ntest_df.head()\n","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"test_df.to_csv('submission.csv', index=False)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"","execution_count":null,"outputs":[]}],"metadata":{"kernelspec":{"language":"python","display_name":"Python 3","name":"python3"},"language_info":{"pygments_lexer":"ipython3","nbconvert_exporter":"python","version":"3.6.4","file_extension":".py","codemirror_mode":{"name":"ipython","version":3},"name":"python","mimetype":"text/x-python"}},"nbformat":4,"nbformat_minor":4}