{"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
\n
The numpy files dataset used in this notebook can be found Here
"},{"metadata":{},"cell_type":"markdown","source":"

Updates:

\n"},{"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\nfrom sklearn.model_selection import KFold\n\nimport warnings\nwarnings.filterwarnings(\"ignore\")","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"def seedAll(seed=355):\n os.environ[\"PYTHONHASHSEED\"] = str(seed)\n np.random.seed(seed)\n tf.random.set_seed(seed)\n random.seed(seed)\nseedAll()","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":"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":"train.head()","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":{"trusted":true},"cell_type":"code","source":"showXray('../input/vinbigdata-chest-xray-abnormalities-detection/train/0108949daa13dc94634a7d650a05c0bb.dicom',train,with_boxes=True)","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(train.shape[0]))\nprint(\"Number of Unique images in train set: {}\".format(train.image_id.nunique()))\nprint(\"Number of Classes: {}\\n\".format(train.class_name.nunique()))\nprint(\"Class Names: {}\".format(list(train.class_name.unique())))","execution_count":null,"outputs":[]},{"metadata":{"_kg_hide-input":true,"trusted":true},"cell_type":"code","source":"print(\"Null Values:\")\ntrain.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(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(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(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= 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":"def generateFolds(n_splits = None):\n kf = KFold(n_splits= n_splits)\n for id,(tr_,val_) in enumerate(kf.split(train_df[\"image_id\"],train_df[\"label\"])):\n train_df.loc[val_,'kfold'] = int(id)\n train_df[\"kfold\"].astype(int)\n\ngenerateFolds(n_splits=5)","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"class DataLoader:\n def __init__(self,path = None,train_df=train_df,val_df=None):\n self.path = path\n self.df = train_df\n self.val_df = val_df\n self.train_list = [f'{img}.npy' for img in train_df[\"image_id\"].unique()]\n np.random.shuffle(self.train_list)\n self.test_list = [f'{img}.npy' for img in val_df[\"image_id\"].unique()]\n np.random.shuffle(self.test_list)\n \n def read_image(self):\n for img in self.train_list:\n im_name = img.split('.npy')[0]\n image = np.load(self.path+img)\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\n \n def getVal(self):\n images = []\n c_labels = []\n bb_labels = []\n for img in self.test_list:\n im_name = img.split('.npy')[0]\n image = np.load(self.path+img)\n temp = self.val_df[self.val_df.image_id==im_name]\n c_label,bb = temp.iloc[0,1],temp.iloc[0,2:].values.astype('float')\n images.append(image)\n c_labels.append(c_label)\n bb_labels.append(bb)\n return np.stack(images,axis=0),(np.stack(c_labels,axis=0),np.stack(bb_labels,axis=0))\n ","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"

Model Building and Training

     "},{"metadata":{"trusted":true,"_kg_hide-input":false,"_kg_hide-output":false},"cell_type":"code","source":"def build():\n in1 = L.Input(shape=(256,256,1))\n \n out1 = L.Conv2D(32,(3,3),activation=\"relu\")(in1)\n out1 = L.Conv2D(32,(3,3),activation=\"relu\")(out1)\n out1 = L.MaxPooling2D((2,2))(out1)\n \n out1 = L.Conv2D(64,(3,3),activation=\"relu\")(out1)\n out1 = L.Conv2D(64,(3,3),activation=\"relu\")(out1)\n out1 = L.MaxPooling2D((2,2))(out1)\n \n out1 = L.Conv2D(128,(3,3),activation=\"relu\")(out1)\n out1 = L.Conv2D(128,(3,3),activation=\"relu\")(out1)\n out1 = L.MaxPooling2D((2,2))(out1)\n out1 = L.Flatten()(out1)\n \n out2 = L.Dense(50,activation=\"relu\",kernel_initializer=\"lecun_normal\")(out1)\n out2 = L.Dense(30,activation=\"relu\",kernel_initializer=\"lecun_normal\")(out2)\n out2 = L.Dense(15,activation=\"sigmoid\",kernel_initializer=\"lecun_normal\",name='class_out')(out2)\n \n out3 = L.Dense(50,activation=\"relu\",kernel_initializer=\"lecun_normal\")(out1)\n out3 = L.Dense(30,activation=\"relu\",kernel_initializer=\"lecun_normal\")(out3)\n out3 = L.Dense(56,activation=\"relu\",kernel_initializer=\"lecun_normal\",name=\"bb_out\")(out3)\n \n model = tf.keras.Model(inputs=in1,outputs=[out2,out3])\n model.compile(loss={'class_out':'categorical_crossentropy','bb_out':'mse'},optimizer=\"adam\")\n return model","execution_count":null,"outputs":[]},{"metadata":{"_kg_hide-input":false,"_kg_hide-output":false,"trusted":true},"cell_type":"code","source":"model = build()","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"tf.keras.utils.plot_model(model)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"

Training Loop

"},{"metadata":{"trusted":true},"cell_type":"code","source":"def getTest(path=None):\n images = []\n for img in tqdm(os.listdir(path)):\n im_name = img.split('.npy')[0]\n image = np.load(path+img)\n images.append(image)\n return np.stack(images,axis=0)\n\nX_test = getTest('../input/xraynumpy/images/test/')","execution_count":null,"outputs":[]},{"metadata":{"trusted":true},"cell_type":"code","source":"class_label = np.zeros((len(X_test),15))\nbb_label = np.zeros((len(X_test),56))\n\nfor fold in range(5):\n print(f'\\nFold: {fold}\\n')\n \n X_train = train_df[train_df.kfold!=fold].drop('kfold',axis=1)\n X_val = train_df[train_df.kfold==fold].drop('kfold',axis=1)\n \n dl = DataLoader('../input/xraynumpy/images/train/',X_train,X_val)\n train_set = dl.flow(batch_size=32)\n X_eval,Y_eval = dl.getVal()\n \n chckpt = tf.keras.callbacks.ModelCheckpoint(f'./model_f{fold}.hdf5',monitor='val_loss',mode='min',save_best_only=True)\n \n K.clear_session()\n model = build()\n \n model.fit(train_set,\n epochs=10,\n steps_per_epoch=int(15000/32),\n validation_data = (X_eval,Y_eval),\n callbacks = [chckpt]\n )\n \n c,b = model.predict(X_test)\n class_label+=c\n bb_label+=b\nclass_label = class_label/5\nbb_label = bb_label/5\nnp.save('./class_label.npy',class_label)\nnp.save('./bb_label.npy',bb_label)","execution_count":null,"outputs":[]},{"metadata":{},"cell_type":"markdown","source":"# Work in Progress....\n

To Do:

\n"},{"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}