{"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_minor":4,"nbformat":4,"cells":[{"cell_type":"markdown","source":"
\n

Diagnosis of breast cancer-Model Development 

\n
","metadata":{}},{"cell_type":"markdown","source":"
\n\n
\n","metadata":{}},{"cell_type":"markdown","source":"# Description\nIn this notebook, me and my team build a **CustomDataset**, a **Dataloader** and a **Deep learning** model. This is part of a problem-solving process that uses deep learning to diagnose breast cancer.\n\n### **Data**\n\n- #### To solve this problem, in addition to RSNA's image data, we also use 3 more datasets: INbreast, MIAS, and DDSM.\n\n### **Model** \n\n- #### In this notebook, we use a basic CNN model combined with fully connected. Detailed model pictures are below.\n\n
\n\n
\n\n\n### Reference\n\nThe architecture mentioned above is the one we refer to [here](https://s3.us-west-2.amazonaws.com/secure.notion-static.com/502af837-13ee-4354-b419-982540af6598/1-s2.0-S0378475422002506-main_%281%29.pdf?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Content-Sha256=UNSIGNED-PAYLOAD&X-Amz-Credential=AKIAT73L2G45EIPT3X45%2F20230120%2Fus-west-2%2Fs3%2Faws4_request&X-Amz-Date=20230120T160710Z&X-Amz-Expires=86400&X-Amz-Signature=df495c3bce0ed9e515bf12a1dae101aa77eb70feacf6127b38c7a9967bbab052&X-Amz-SignedHeaders=host&response-content-disposition=filename%3D%221-s2.0-S0378475422002506-main%2520%281%29.pdf%22&x-id=GetObject).\n\n\n\n\n","metadata":{}},{"cell_type":"markdown","source":"# 1. Install lib","metadata":{}},{"cell_type":"code","source":"!pip install /kaggle/input/rsnawhl/{pydicom-2.3.0-py3-none-any.whl,pylibjpeg-1.4.0-py3-none-any.whl,python_gdcm-3.0.15-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl}","metadata":{"execution":{"iopub.status.busy":"2023-02-06T13:54:53.747214Z","iopub.execute_input":"2023-02-06T13:54:53.747581Z","iopub.status.idle":"2023-02-06T13:55:26.7425Z","shell.execute_reply.started":"2023-02-06T13:54:53.747549Z","shell.execute_reply":"2023-02-06T13:55:26.741313Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"!pip install /kaggle/input/nvidia-dali-wheel/nvidia_dali_nightly_cuda110-1.22.0.dev20221213-6757685-py3-none-manylinux2014_x86_64.whl\n!pip install /kaggle/input/nvidia-dali-wheel/dicomsdl-0.109.1-cp37-cp37m-manylinux_2_12_x86_64.manylinux2010_x86_64.whl","metadata":{"execution":{"iopub.status.busy":"2023-02-06T13:55:26.745133Z","iopub.execute_input":"2023-02-06T13:55:26.745619Z","iopub.status.idle":"2023-02-06T13:56:39.88656Z","shell.execute_reply.started":"2023-02-06T13:55:26.745555Z","shell.execute_reply":"2023-02-06T13:56:39.885371Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# pip install torchsummary wandb pydicom","metadata":{"jupyter":{"source_hidden":true},"execution":{"iopub.status.busy":"2023-02-06T13:56:39.888825Z","iopub.execute_input":"2023-02-06T13:56:39.889976Z","iopub.status.idle":"2023-02-06T13:56:39.895414Z","shell.execute_reply.started":"2023-02-06T13:56:39.889932Z","shell.execute_reply":"2023-02-06T13:56:39.894387Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# pip install dicomsdl","metadata":{"jupyter":{"source_hidden":true},"execution":{"iopub.status.busy":"2023-02-06T13:56:39.898427Z","iopub.execute_input":"2023-02-06T13:56:39.898866Z","iopub.status.idle":"2023-02-06T13:56:39.909586Z","shell.execute_reply.started":"2023-02-06T13:56:39.898832Z","shell.execute_reply":"2023-02-06T13:56:39.908633Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# 2. Import lib","metadata":{}},{"cell_type":"code","source":"import cv2 \nimport torch\nimport imghdr\nimport pydicom\nimport plistlib\nimport numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\n\nimport dicomsdl\nimport torch.nn as nn\nfrom PIL import Image\nimport os, sys, tarfile\nfrom skimage import filters\nfrom skimage.io import imread\nimport pytorch_lightning as pl\nfrom tqdm.notebook import tqdm\nimport torch.nn.functional as F\n# from torchsummary import summary\nfrom skimage.draw import polygon\nimport torchvision.transforms as T\nfrom torchvision.io import read_image\nfrom skimage.exposure import equalize_adapthist\nfrom torch.utils.data import DataLoader, Dataset, ConcatDataset\n\n# Data Augmentation for Image Preprocessing\nfrom albumentations.pytorch import ToTensorV2\nfrom albumentations import (ToFloat, Normalize, VerticalFlip, HorizontalFlip, Compose, Resize,\n RandomBrightnessContrast, HueSaturationValue, Blur, GaussNoise,\n Rotate, RandomResizedCrop, Cutout, ShiftScaleRotate, ToGray)\n\n%matplotlib inline","metadata":{"execution":{"iopub.status.busy":"2023-02-06T13:56:39.912874Z","iopub.execute_input":"2023-02-06T13:56:39.913142Z","iopub.status.idle":"2023-02-06T13:56:45.929263Z","shell.execute_reply.started":"2023-02-06T13:56:39.913119Z","shell.execute_reply":"2023-02-06T13:56:45.928282Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# 3. Dataset and Dataloader","metadata":{}},{"cell_type":"markdown","source":"### INBREAST DATASET","metadata":{}},{"cell_type":"code","source":"class INBreastDataset(Dataset):\n\n '''\n INBreast dataset\n '''\n def __init__(self, csv_file: str, \n root_dir: str, \n transform=None):\n '''\n Args:\n csv_file (string): Path to the csv file with annotations.\n root_dir (string): Directory with all the images.\n transform (callable, optional): Optional transform to be applied\n on a sample.\n '''\n self.img_dir = csv_file\n self.img_data = pd.read_csv(csv_file)\n self.root = root_dir\n self.transform = transform\n \n def __len__(self):\n return len(self.img_data['Lesion Annotation Status'])\n \n def pre_img(self, img):\n img_clahe = equalize_adapthist(img, clip_limit=0.04, nbins=256)\n return img_clahe\n\n def __getitem__(self, idx):\n img_path = os.path.join(self.root, self.img_data.iloc[idx, -1].replace('dcm', 'png'))\n img = cv2.imread(img_path, 0)\n img = self.pre_img(img)\n if self.transform != None:\n image_trans = self.transform(image=img)['image']\n else:\n image_trans = img\n #return\n final_image = image_trans\n label = self.img_data['Lesion Annotation Status'][idx]\n return final_image, label","metadata":{"jupyter":{"source_hidden":true},"execution":{"iopub.status.busy":"2023-02-06T13:56:45.930734Z","iopub.execute_input":"2023-02-06T13:56:45.931576Z","iopub.status.idle":"2023-02-06T13:56:45.942881Z","shell.execute_reply.started":"2023-02-06T13:56:45.931536Z","shell.execute_reply":"2023-02-06T13:56:45.94135Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"### RSNA DATASET","metadata":{}},{"cell_type":"code","source":"class RSNADataset(Dataset):\n \n '''\n RSNA Dataset\n '''\n def __init__(self,csv_file: str,\n root_dir : str,\n transform=None,is_train = True):\n '''\n Args:\n csv_file (string): Path to the csv file with annotations.\n root_dir (string): Directory with all the images.\n transform (callable, optional): Optional transform to be applied\n on a sample.\n '''\n self.dataframe = pd.read_csv(csv_file)\n self.root_dir = root_dir\n self.is_train = is_train\n \n if is_train:\n self.transform = transform\n else:\n self.transform = Compose([Resize(height=227,width=227,always_apply=True),\n ToTensorV2()])\n \n def __len__(self):\n return len(self.dataframe)\n \n \n def get_cdf(self, img):\n m = int(np.max(img))\n hist = np.histogram(img, bins=m+1, range=(0, m+1))[0]\n hist = hist/img.size\n cdf = np.cumsum(hist)\n return cdf\n\n # numpy\n def histogram_equalization(self, img):\n m = int(np.max(img))\n hist = np.histogram(img, bins=m+1, range=(0, m+1))[0]\n # bước 1: tính pdf\n hist = hist/img.size\n # bước 2: tính cdf\n cdf = np.cumsum(hist)\n # bước 3: lập bảng thay thế\n s_k = (255 * cdf)\n # ảnh mới\n img_new = np.array([s_k[i] for i in img.ravel()]).reshape(img.shape)\n return img_new\n\n\n def HE(self, img):\n '''\n input: gray image\n dtype: np.array\n '''\n img_equ= self.histogram_equalization(img)\n return img_equ\n\n def pre_img(self, img):\n# clahe = cv2.createCLAHE(clipLimit = 20)\n# img_clahe = clahe.apply(img) + 30\n img_clahe = equalize_adapthist(img, clip_limit=0.04, nbins=256)\n# HE_img = self.HE(img_clahe)\n# print(HE_img)\n return img_clahe\n\n def __getitem__(self,index):\n if self.is_train:\n if self.dataframe.iloc[index]['cancer']==0:\n image_path = self.root_dir + '/nocancer/'+str(self.dataframe.iloc[index].patient_id) + \"_\" + str(self.dataframe.iloc[index].image_id) + \".png\"\n else:\n image_path = self.root_dir + '/cancer/'+str(self.dataframe.iloc[index].patient_id) + \"_\" + str(self.dataframe.iloc[index].image_id) + \".png\"\n else:\n image_path = self.root_dir +'/'+ str(self.dataframe.iloc[index].patient_id) + \"/\" + str(self.dataframe.iloc[index].image_id) + \".png\"\n image = cv2.imread(image_path,0)\n image = self.pre_img(image)\n \n #transform image\n # return a dict have key 'image'\n if self.transform != None:\n image_trans = self.transform(image=image)['image']\n else:\n image_trans = image\n \n #return\n final_image = image_trans\n if self.is_train:\n label = self.dataframe.iloc[index]['cancer'] \n return final_image,label\n else: \n return final_image","metadata":{"jupyter":{"source_hidden":true},"execution":{"iopub.status.busy":"2023-02-06T13:56:45.944296Z","iopub.execute_input":"2023-02-06T13:56:45.947551Z","iopub.status.idle":"2023-02-06T13:56:45.963463Z","shell.execute_reply.started":"2023-02-06T13:56:45.947515Z","shell.execute_reply":"2023-02-06T13:56:45.962641Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"### MIAS DATASET","metadata":{}},{"cell_type":"code","source":"class MIASDataset(Dataset):\n def __init__(self, \n csv_file : str,\n root: str,\n transform=None):\n self.img_dir = root\n self.total_imgs = pd.read_csv(csv_file)\n self.transform = transform\n\n def __len__(self):\n return len(self.total_imgs)\n\n def pre_img(self, img):\n img_clahe = equalize_adapthist(img, clip_limit=0.04, nbins=256)\n return img_clahe\n \n def __getitem__(self, idx):\n img_path = self.total_imgs['Images path'][idx]\n image = cv2.imread(img_path, 0)\n label = self.total_imgs['Labels'][idx]\n image = self.pre_img(image)\n if self.transform != None:\n image_trans = self.transform(image=image)['image']\n else:\n image_trans = image\n #return\n final_image = image_trans\n# print(image_trans.shape)\n return final_image, label","metadata":{"jupyter":{"source_hidden":true},"execution":{"iopub.status.busy":"2023-02-06T13:56:45.96471Z","iopub.execute_input":"2023-02-06T13:56:45.965134Z","iopub.status.idle":"2023-02-06T13:56:45.979513Z","shell.execute_reply.started":"2023-02-06T13:56:45.965094Z","shell.execute_reply":"2023-02-06T13:56:45.978652Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"class clr:\n START = '\\033[1m'\n PURPLE = '\\033[95m'\n CYAN = '\\033[96m'\n DARKCYAN = '\\033[36m'\n BLUE = '\\033[94m'\n GREEN = '\\033[92m'\n YELLOW = '\\033[93m'\n RED = '\\033[91m'\n BOLD = '\\033[1m'\n UNDERLINE = '\\033[4m'\n END = '\\033[0m'\n","metadata":{"jupyter":{"source_hidden":true},"execution":{"iopub.status.busy":"2023-02-06T13:56:45.980945Z","iopub.execute_input":"2023-02-06T13:56:45.981362Z","iopub.status.idle":"2023-02-06T13:56:45.991817Z","shell.execute_reply.started":"2023-02-06T13:56:45.981322Z","shell.execute_reply":"2023-02-06T13:56:45.990591Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"## Building Datamodule ","metadata":{}},{"cell_type":"code","source":"class Datamodule(pl.LightningDataModule):\n \n '''\n Datamodule to load data\n '''\n def __init__(self, \n batch_size: int,\n train_set,\n test_set=None,\n validate_set=None):\n super().__init__()\n\n self.train_set = train_set\n self.test_set = test_set\n self.validate_set = validate_set\n self.batch_size = batch_size\n\n\n def train_dataloader(self):\n '''\n DataLoader of train func\n '''\n return DataLoader(\n self.train_set,\n batch_size=self.batch_size,\n shuffle=True, # shuffle dữ liệu để tạo sự đa dạng,\n drop_last=True\n )\n\n def val_dataloader(self):\n '''\n DataLoader of validate func\n '''\n return DataLoader(\n self.validate_set,\n batch_size=self.batch_size,\n )\n\n def test_dataloader(self):\n '''\n DataLoader of test func\n '''\n return DataLoader(\n self.test_set,\n batch_size=self.batch_size,\n )\n","metadata":{"jupyter":{"source_hidden":true},"execution":{"iopub.status.busy":"2023-02-06T13:56:45.998162Z","iopub.execute_input":"2023-02-06T13:56:45.999036Z","iopub.status.idle":"2023-02-06T13:56:46.010839Z","shell.execute_reply.started":"2023-02-06T13:56:45.999002Z","shell.execute_reply":"2023-02-06T13:56:46.009339Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# 4. Model","metadata":{}},{"cell_type":"markdown","source":"### a. Model base","metadata":{}},{"cell_type":"code","source":"class CNN(pl.LightningModule):\n def __init__(self,\n width:int=227,\n height:int=227):\n super().__init__()\n def layer(input_channel: int, \n output_channel: int, \n kernel_size_conv: int=3, \n padding: str='same', \n stride_conv: int=1, \n kernel_size_maxpool: int=2, \n stride_maxpool: int=2,\n normalize=True,\n maxpool=True):\n layers = [nn.Conv2d(input_channel, \n output_channel, \n kernel_size_conv,\n stride_conv,\n padding\n )]\n if normalize:\n layers.append(nn.BatchNorm2d(output_channel))\n layers.append(nn.ReLU())\n if maxpool:\n layers.append(nn.MaxPool2d(kernel_size=kernel_size_maxpool, \n stride=stride_maxpool))\n return layers\n \n self.model = nn.Sequential(\n *layer(input_channel=1,\n output_channel=16),\n *layer(input_channel=16,\n output_channel=32),\n *layer(input_channel=32,\n output_channel=64,\n maxpool=False)\n )\n \n self.relu = nn.ReLU()\n self.fc1 = nn.Linear((width//4) * (height//4) * 64, 512)\n self.fc2 = nn.Linear(512, 128)\n self.fc3 = nn.Linear(128, 2)\n self.softmax = nn.Softmax(dim=1)\n def forward(self, x):\n# print(type(x))\n x = self.model(x.float())\n x = torch.flatten(x, 1)\n x = self.relu(self.fc1(x))\n x = self.relu(self.fc2(x))\n# print(x)\n x = self.fc3(x) \n x = self.softmax(x)\n return x","metadata":{"execution":{"iopub.status.busy":"2023-02-06T13:56:46.012443Z","iopub.execute_input":"2023-02-06T13:56:46.013165Z","iopub.status.idle":"2023-02-06T13:56:46.025068Z","shell.execute_reply.started":"2023-02-06T13:56:46.013132Z","shell.execute_reply":"2023-02-06T13:56:46.024242Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"### b. Model using pytorchlighning","metadata":{}},{"cell_type":"code","source":"class FullModel(pl.LightningModule):\n def __init__(self, \n lr: float, \n total_steps: int,\n width:int=227,\n height:int=227\n ):\n super().__init__()\n self.lr = lr\n self.model = CNN(width=width, height=height)\n self.criterion = nn.CrossEntropyLoss()\n self.total_steps = total_steps\n self.save_hyperparameters()\n self.dem = 0\n \n def configure_optimizers(self):\n# optimizer = torch.optim.AdamW(self.parameters(), lr=self.lr, betas=(0.9, 0.98), weight_decay=1e-4)\n optimizer = torch.optim.SGD(model.parameters(), lr=0.1)\n# optimizer = torch.optim.RMSprop(self.parameters(), lr=0.0001)\n scheduler = torch.optim.lr_scheduler.OneCycleLR(\n optimizer, max_lr=self.lr, total_steps=self.total_steps, verbose=False,\n )\n scheduler = {\n \"scheduler\": scheduler,\n \"interval\": \"step\", # or 'epoch'\n \"frequency\": 1,\n }\n return [optimizer], [scheduler]\n \n def forward(self, x):\n return self.model(x)\n\n def training_step(self, batch, batch_idx):\n '''\n Hàm này nhận vào một batch\n lấy kết quả của mô hình và tính loss\n mô hình sẽ tự tính back propagation\n -----------------------------------\n This function takes a batch, gets the \n result of the model, and calculates the loss. \n The model will automatically calculate back propagation.\n '''\n img, labels = batch\n# print(img[0].shape, self.dem)\n outputs = self.model(img)\n loss = self.criterion(outputs, labels)\n self.log(\"train_loss\", loss.item())\n #Accuracy\n output = torch.argmax(outputs, dim=1)\n correct = (output == labels).float().sum()\n# print(\"train\", labels, output)\n self.log(\"train_loss\", loss.item())\n self.log(\"train_acc\", correct/len(labels))\n# self.dem+=1\n return loss\n\n def validation_step(self, batch, batch_idx):\n '''\n Hàm này nhận vào một batch\n lấy kết quả của mô hình và tính loss.\n ---------------------------------------------\n This function takes a batch, gets the \n result of the model, and calculates the loss. \n '''\n img, labels = batch\n outputs = self.model(img)\n loss = self.criterion(outputs, labels)\n #Accuracy\n output = torch.argmax(outputs, dim=1)\n# print(\"val\", labels, output)\n correct = (output == labels).float().sum()\n\n # lưu lại loss của validate\n self.log(\"val_loss\", loss.item())\n self.log(\"val_acc\", correct/len(labels))\n\n return loss\n\n def test_step(self, batch, batch_idx):\n img, labels = batch\n outputs = self.model(img)\n loss = self.criterion(outputs, labels)\n\n self.log(\"train_loss\", loss.item())\n\n # lưu lại loss của test\n self.log(\"test_loss\", loss.item())\n self.log(\"test_acc\", acc.item())\n\n return loss\n","metadata":{"execution":{"iopub.status.busy":"2023-02-06T13:56:46.026494Z","iopub.execute_input":"2023-02-06T13:56:46.027115Z","iopub.status.idle":"2023-02-06T13:56:46.047875Z","shell.execute_reply.started":"2023-02-06T13:56:46.027082Z","shell.execute_reply":"2023-02-06T13:56:46.046955Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"# 5. LOGIN WANDB","metadata":{}},{"cell_type":"markdown","source":"#### A tool for visualizing and tracking your machine learning experiments. This repo contains the CLI and Python API.","metadata":{}},{"cell_type":"code","source":"# pip install wandb --upgrade","metadata":{"execution":{"iopub.status.busy":"2023-02-06T13:56:46.049306Z","iopub.execute_input":"2023-02-06T13:56:46.050005Z","iopub.status.idle":"2023-02-06T13:56:46.06273Z","shell.execute_reply.started":"2023-02-06T13:56:46.049972Z","shell.execute_reply":"2023-02-06T13:56:46.061679Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"#### Log in to Wandb if you have an account","metadata":{}},{"cell_type":"code","source":"# from kaggle_secrets import UserSecretsClient\n# user_secrets = UserSecretsClient()\n# secret_value_0 = user_secrets.get_secret(\"wandb\")\n# !wandb login $secret_value_0","metadata":{"execution":{"iopub.status.busy":"2023-02-06T13:56:46.064197Z","iopub.execute_input":"2023-02-06T13:56:46.064931Z","iopub.status.idle":"2023-02-06T13:56:46.073654Z","shell.execute_reply.started":"2023-02-06T13:56:46.064899Z","shell.execute_reply":"2023-02-06T13:56:46.072922Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# pd.read_csv(\"/kaggle/input/breastcancer/RSNA_metadata.csv\")","metadata":{"execution":{"iopub.status.busy":"2023-02-06T13:56:46.075283Z","iopub.execute_input":"2023-02-06T13:56:46.075964Z","iopub.status.idle":"2023-02-06T13:56:46.08434Z","shell.execute_reply.started":"2023-02-06T13:56:46.075921Z","shell.execute_reply":"2023-02-06T13:56:46.083392Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"### a. Reading dataset","metadata":{}},{"cell_type":"code","source":"# root_inbreast = \"/kaggle/input/inbreast-crop-png/ALL-IMGS\"\n# csv_inbreast = \"/kaggle/input/breastcancer/INBreast_config.csv\"\n# csv_RSNA = '/kaggle/input/breastcancer/RSNA_metadata.csv'\n# root_RSNA = '/kaggle/input/rsna-256px-croped/RSNA_Crop'\n\n# root_test_RSNA = '/kaggle/input/test-rsna-dataset'\ncsv_test_RSNA = '/kaggle/input/rsna-breast-cancer-detection/test.csv'\n# root_MIAS='/kaggle/input/mias-png'\n# csv_MIAS= '/kaggle/input/breastcancer/mias_metadata.csv'\n\n# transform = Compose([Resize(height=227,width=227,always_apply=True),\n# # Normalize(mean=0.449,std=0.226),\n# HorizontalFlip(),\n# VerticalFlip(),\n# Rotate(),\n# ToTensorV2()])\n\n# train_inbreast = INBreastDataset(csv_inbreast, root_inbreast, transform)\n# train_rsna = RSNADataset(csv_RSNA,root_RSNA,transform)\n# train_mias = MIASDataset(csv_MIAS, root_MIAS, transform)\n# test_rsna = RSNADataset(csv_test_RSNA, root_test_RSNA, transform=transform, is_train=False)\n# print(len(train_inbreast), len(train_rsna), len(test_rsna))","metadata":{"execution":{"iopub.status.busy":"2023-02-06T13:56:46.085851Z","iopub.execute_input":"2023-02-06T13:56:46.086515Z","iopub.status.idle":"2023-02-06T13:56:46.096443Z","shell.execute_reply.started":"2023-02-06T13:56:46.08648Z","shell.execute_reply":"2023-02-06T13:56:46.095457Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# mias = pd.read_csv(csv_MIAS)\n# inbrest = pd.read_csv(csv_inbreast)\n# rsna = pd.read_csv(csv_RSNA)\n# np.unique(inbrest['Lesion Annotation Status'], return_counts=True)\n","metadata":{"jupyter":{"source_hidden":true},"execution":{"iopub.status.busy":"2023-02-06T13:56:46.098021Z","iopub.execute_input":"2023-02-06T13:56:46.098721Z","iopub.status.idle":"2023-02-06T13:56:46.111519Z","shell.execute_reply.started":"2023-02-06T13:56:46.098688Z","shell.execute_reply":"2023-02-06T13:56:46.110384Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# np.unique(rsna.cancer, return_counts=True)","metadata":{"jupyter":{"source_hidden":true},"execution":{"iopub.status.busy":"2023-02-06T13:56:46.113052Z","iopub.execute_input":"2023-02-06T13:56:46.11366Z","iopub.status.idle":"2023-02-06T13:56:46.121902Z","shell.execute_reply.started":"2023-02-06T13:56:46.113625Z","shell.execute_reply":"2023-02-06T13:56:46.120894Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# np.unique(mias.Labels, return_counts=True)","metadata":{"jupyter":{"source_hidden":true},"execution":{"iopub.status.busy":"2023-02-06T13:56:46.123477Z","iopub.execute_input":"2023-02-06T13:56:46.124101Z","iopub.status.idle":"2023-02-06T13:56:46.13181Z","shell.execute_reply.started":"2023-02-06T13:56:46.124068Z","shell.execute_reply":"2023-02-06T13:56:46.130855Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# train_dataset = ConcatDataset([train_inbreast, train_mias])\n# test_dataset = ConcatDataset([test_rsna])\n# dm = Datamodule(32, train_dataset, train_rsna, train_rsna)","metadata":{"jupyter":{"source_hidden":true},"execution":{"iopub.status.busy":"2023-02-06T13:56:46.133387Z","iopub.execute_input":"2023-02-06T13:56:46.134069Z","iopub.status.idle":"2023-02-06T13:56:46.142644Z","shell.execute_reply.started":"2023-02-06T13:56:46.134035Z","shell.execute_reply":"2023-02-06T13:56:46.141541Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"### b. Visualize the image after preprocessing.","metadata":{}},{"cell_type":"code","source":"def visulize(dataset, mode='train'):\n plt.figure(figsize=(15, 15))\n if mode == 'train':\n for i, (img, label) in enumerate(dataset):\n # print(img.permute(1, 2, 0).shape, label)\n # print(img.shape)\n plt.subplot(1,10,i+1)\n plt.imshow(img.squeeze())\n plt.axis('off')\n plt.subplots_adjust(wspace=None, hspace=None)\n plt.title(label)\n if i == 9:\n break\n else:\n for i, (img) in enumerate(dataset):\n # print(img.permute(1, 2, 0).shape, label)\n # print(img.shape)\n plt.subplot(1,4,i+1)\n plt.imshow(img.squeeze())\n plt.axis('off')\n plt.subplots_adjust(wspace=None, hspace=None)\n if i == 3:\n break\n# visulize(train_dataset, mode='train')","metadata":{"jupyter":{"source_hidden":true},"execution":{"iopub.status.busy":"2023-02-06T13:56:46.144434Z","iopub.execute_input":"2023-02-06T13:56:46.145091Z","iopub.status.idle":"2023-02-06T13:56:46.153857Z","shell.execute_reply.started":"2023-02-06T13:56:46.145059Z","shell.execute_reply":"2023-02-06T13:56:46.152891Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# visulize(test_dataset, mode='test')","metadata":{"execution":{"iopub.status.busy":"2023-02-06T13:56:46.155423Z","iopub.execute_input":"2023-02-06T13:56:46.156106Z","iopub.status.idle":"2023-02-06T13:56:46.169414Z","shell.execute_reply.started":"2023-02-06T13:56:46.156072Z","shell.execute_reply":"2023-02-06T13:56:46.168478Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"### c. Create model and training","metadata":{}},{"cell_type":"code","source":"max_epochs = 100\nlr = 0.05\n# total_step = len(dm.train_dataloader()) * max_epochs\n# model = FullModel(lr, \n# 0,\n# width=227,\n# height=227)","metadata":{"execution":{"iopub.status.busy":"2023-02-06T13:56:46.170948Z","iopub.execute_input":"2023-02-06T13:56:46.171556Z","iopub.status.idle":"2023-02-06T13:56:46.180621Z","shell.execute_reply.started":"2023-02-06T13:56:46.171523Z","shell.execute_reply":"2023-02-06T13:56:46.179655Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# wandb_logger = pl.loggers.WandbLogger(\n# project=\"BREAST-CANCER\", name='model_CNN_ver3', log_model=\"all\"\n# )\n# lr_monitor = pl.callbacks.LearningRateMonitor(logging_interval=\"step\")\n\n# trainer = pl.Trainer(\n# logger=wandb_logger,\n# callbacks=[lr_monitor],\n# max_epochs=max_epochs,\n# accelerator=\"auto\",\n# log_every_n_steps=1\n# gradient_clip_val=4\n# )\n\n# trainer.fit(model, datamodule=dm)#, ckpt_path='/kaggle/working/depth_estimation/26ozxilx/checkpoints/epoch=1-step=2850.ckpt')","metadata":{"execution":{"iopub.status.busy":"2023-02-06T13:56:46.182305Z","iopub.execute_input":"2023-02-06T13:56:46.183073Z","iopub.status.idle":"2023-02-06T13:56:46.19201Z","shell.execute_reply.started":"2023-02-06T13:56:46.183041Z","shell.execute_reply":"2023-02-06T13:56:46.191102Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"### d. Save checkpoint","metadata":{}},{"cell_type":"code","source":"# trainer.save_checkpoint(\"best_model.ckpt\")","metadata":{"execution":{"iopub.status.busy":"2023-02-06T13:56:46.193594Z","iopub.execute_input":"2023-02-06T13:56:46.194291Z","iopub.status.idle":"2023-02-06T13:56:46.202909Z","shell.execute_reply.started":"2023-02-06T13:56:46.194258Z","shell.execute_reply":"2023-02-06T13:56:46.201673Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"### e. Predict","metadata":{}},{"cell_type":"code","source":"PATH = '/kaggle/input/ckpt-rsna/model_ver3.1.ckpt'\n\nmodel = FullModel(lr, \n 0,\n width=227,\n height=227).load_from_checkpoint(PATH)","metadata":{"execution":{"iopub.status.busy":"2023-02-06T13:56:46.204404Z","iopub.execute_input":"2023-02-06T13:56:46.205008Z","iopub.status.idle":"2023-02-06T13:56:59.361993Z","shell.execute_reply.started":"2023-02-06T13:56:46.204975Z","shell.execute_reply":"2023-02-06T13:56:59.360915Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# model.eval()","metadata":{"execution":{"iopub.status.busy":"2023-02-06T13:56:59.363541Z","iopub.execute_input":"2023-02-06T13:56:59.364226Z","iopub.status.idle":"2023-02-06T13:56:59.369063Z","shell.execute_reply.started":"2023-02-06T13:56:59.364187Z","shell.execute_reply":"2023-02-06T13:56:59.368108Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# trainer.validate(model, dataloaders=dm)","metadata":{"jupyter":{"source_hidden":true},"execution":{"iopub.status.busy":"2023-02-06T13:56:59.37604Z","iopub.execute_input":"2023-02-06T13:56:59.376804Z","iopub.status.idle":"2023-02-06T13:56:59.383541Z","shell.execute_reply.started":"2023-02-06T13:56:59.376765Z","shell.execute_reply":"2023-02-06T13:56:59.382423Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# def pre_img(img):\n# img_clahe = equalize_adapthist(img, clip_limit=0.04, nbins=256)\n# return img_clahe\n# # img = cv2.imread(\"/kaggle/input/rsna-256px-croped/RSNA_Crop/cancer/10130_1360338805.png\", 0)\n# img = cv2.imread(\"/kaggle/input/test-rsna-dataset/10008/1591370361.png\", 0)\n# img = pre_img(img)\n# transform = Compose([Resize(height=227,width=227,always_apply=True),\n# ToTensorV2()])\n# img = transform(image=img)['image']\n# # plt.imshow(img.squeeze())","metadata":{"jupyter":{"source_hidden":true},"execution":{"iopub.status.busy":"2023-02-06T13:56:59.386998Z","iopub.execute_input":"2023-02-06T13:56:59.387317Z","iopub.status.idle":"2023-02-06T13:56:59.396497Z","shell.execute_reply.started":"2023-02-06T13:56:59.387283Z","shell.execute_reply":"2023-02-06T13:56:59.395468Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# visulize(test_dataset, mode='test')","metadata":{"jupyter":{"source_hidden":true},"execution":{"iopub.status.busy":"2023-02-06T13:56:59.398025Z","iopub.execute_input":"2023-02-06T13:56:59.39843Z","iopub.status.idle":"2023-02-06T13:56:59.411643Z","shell.execute_reply.started":"2023-02-06T13:56:59.398393Z","shell.execute_reply":"2023-02-06T13:56:59.410734Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# lst_test = !ls /kaggle/input/test-rsna-dataset/10008\n# for i in lst_test:\n# img = cv2.imread(\"/kaggle/input/test-rsna-dataset/10008/\"+i, 0)\n# img = pre_img(img)\n# transform = Compose([Resize(height=227,width=227,always_apply=True),\n# ToTensorV2()])\n# img = transform(image=img)['image']\n# print(model(img.unsqueeze(0)).argmax().item())","metadata":{"jupyter":{"source_hidden":true},"execution":{"iopub.status.busy":"2023-02-06T13:56:59.413158Z","iopub.execute_input":"2023-02-06T13:56:59.413574Z","iopub.status.idle":"2023-02-06T13:56:59.422343Z","shell.execute_reply.started":"2023-02-06T13:56:59.413536Z","shell.execute_reply":"2023-02-06T13:56:59.421449Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"### f. Save file predict to submit\n","metadata":{}},{"cell_type":"code","source":"# df = pd.read_csv(csv_test_RSNA)\n# lst_test = ['/kaggle/input/rsna-breast-cancer-detection/test_images/' + str(df.patient_id[i]) + '/' + str(df.image_id[i]) + '.dcm' for i in range(len(df))]\n","metadata":{"jupyter":{"source_hidden":true},"execution":{"iopub.status.busy":"2023-02-06T13:56:59.424093Z","iopub.execute_input":"2023-02-06T13:56:59.42445Z","iopub.status.idle":"2023-02-06T13:56:59.433161Z","shell.execute_reply.started":"2023-02-06T13:56:59.424416Z","shell.execute_reply":"2023-02-06T13:56:59.432229Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# checkpoint=torch.load(PATH)\n# print(checkpoint['state_dict'])","metadata":{"jupyter":{"source_hidden":true},"execution":{"iopub.status.busy":"2023-02-06T13:56:59.434816Z","iopub.execute_input":"2023-02-06T13:56:59.435179Z","iopub.status.idle":"2023-02-06T13:56:59.444146Z","shell.execute_reply.started":"2023-02-06T13:56:59.435146Z","shell.execute_reply":"2023-02-06T13:56:59.443199Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def pre_img(img):\n img_clahe = equalize_adapthist(img, clip_limit=0.04, nbins=256)\n return img_clahe\ndef ROI(img):\n # smooth image\n blur = cv2.GaussianBlur(img,(5, 5), 0)\n # threshold image\n threshold = ((blur > np.mean(img))*255).astype(np.uint8)\n # Apply the Component analysis function\n analysis = cv2.connectedComponentsWithStats(threshold, 4, cv2.CV_32S)\n (totalLabels, label_ids, values, centroid) = analysis\n # Loop through each component\n crop_img = None\n for i in range(1, totalLabels):\n area = values[i, cv2.CC_STAT_AREA] \n s = img.size*0.1\n if (area > s):\n # Create a new image for bounding boxes\n new_img = img.copy()\n # Now extract the coordinate points\n x = values[i, cv2.CC_STAT_LEFT]\n y = values[i, cv2.CC_STAT_TOP]\n w = values[i, cv2.CC_STAT_WIDTH]\n h = values[i, cv2.CC_STAT_HEIGHT]\n crop_img = new_img[y:y+h, x:x+w]\n return crop_img\n \ndef make_predict_file(csv_path, model):\n df = pd.read_csv(csv_path)\n df['prediction_id'] = [str(df.patient_id[i]) + '_' + str(df.laterality[i]) for i in range(len(df))]\n if(csv_path.split('/')[-1] =='test.csv'):\n lst_test = ['/kaggle/input/rsna-breast-cancer-detection/test_images/' + str(df.patient_id[i]) + '/' + str(df.image_id[i]) + '.dcm' for i in range(len(df))]\n else:\n lst_test = ['/kaggle/input/rsna-breast-cancer-detection/train_images/' + str(df.patient_id[i]) + '/' + str(df.image_id[i]) + '.dcm' for i in range(len(df))]\n lst_predict = []\n for path in tqdm(lst_test):\n # read describe image\n# dcm = pydicom.dcmread(path)\n try:\n dcm = dicomsdl.open(path)\n img = dcm.pixelData(storedvalue=False) # storedvalue = True for int16 return otherwise float32\n # img = cv2.imread(path, 0)\n # read image\n # img = dcm.pixel_array\n img = ((img-img.min())/(img.max()-img.min()))*255\n # invert pixel\n if dcm.PhotometricInterpretation=='MONOCHROME1':\n img = img.max()-img\n\n img = ROI(img)\n img = img.astype('uint8')\n img = cv2.resize(img, (227, 227))\n img = pre_img(img)\n # plt.imshow(img)\n transform = Compose([Resize(height=227,width=227,always_apply=True),\n ToTensorV2()])\n img = transform(image=img)['image']\n lst_predict.append(model(img.unsqueeze(0)).argmax().item())\n except:\n pass\n submission = pd.DataFrame({'prediction_id': df.prediction_id, 'cancer':lst_predict}).groupby(\"prediction_id\").mean().reset_index()\n return submission","metadata":{"execution":{"iopub.status.busy":"2023-02-06T13:56:59.445651Z","iopub.execute_input":"2023-02-06T13:56:59.446035Z","iopub.status.idle":"2023-02-06T13:56:59.511592Z","shell.execute_reply.started":"2023-02-06T13:56:59.445988Z","shell.execute_reply":"2023-02-06T13:56:59.510584Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"path = '/kaggle/input/rsna-breast-cancer-detection/train.csv'\nmake_predict_file(csv_test_RSNA, model).to_csv(\"submission.csv\", index=False)","metadata":{"execution":{"iopub.status.busy":"2023-02-06T13:56:59.513158Z","iopub.execute_input":"2023-02-06T13:56:59.513822Z","iopub.status.idle":"2023-02-06T13:57:02.950255Z","shell.execute_reply.started":"2023-02-06T13:56:59.513783Z","shell.execute_reply":"2023-02-06T13:57:02.949277Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# make_predict_file(path, model)","metadata":{"execution":{"iopub.status.busy":"2023-02-06T13:57:02.951908Z","iopub.execute_input":"2023-02-06T13:57:02.952637Z","iopub.status.idle":"2023-02-06T13:57:02.957793Z","shell.execute_reply.started":"2023-02-06T13:57:02.952582Z","shell.execute_reply":"2023-02-06T13:57:02.956424Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"","metadata":{},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"","metadata":{},"execution_count":null,"outputs":[]}]}