{"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":"

Image Processing - Resize and Augmentation

\n\n\n
    \n
  1. Data Description
    2. \n
  2. Load Original DICOM Images
    3. \n
      \n
    1. Cancer Positive Images
      2. \n
    2. Cancer Negative Images
      3. \n
    3. Unusually Difficult Cases
      4. \n
    \n
  3. Resize images using PIL package
    4. \n
  4. Image Augmentation by torchvisaion tranforms
    5. \n
      \n
    1. 4.1. center crop the image and pad image
      2. \n
    2. 4.2. Add Gaussian Blur to images
      3. \n
    \n \n \n
\n\n\nDICOM: Digital Imaging and Communications in Medicine\n\n","metadata":{}},{"cell_type":"code","source":"!pip install -qU python-gdcm pydicom pylibjpeg","metadata":{"execution":{"iopub.status.busy":"2022-12-02T14:02:48.33587Z","iopub.execute_input":"2022-12-02T14:02:48.336842Z","iopub.status.idle":"2022-12-02T14:03:05.989262Z","shell.execute_reply.started":"2022-12-02T14:02:48.336712Z","shell.execute_reply":"2022-12-02T14:03:05.987493Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"import os\nimport cv2\nimport glob\nimport gdcm\nimport pydicom\nimport numpy as np\nimport pandas as pd\nimport seaborn as sns\nimport matplotlib.pyplot as plt\nimport seaborn as sns\n\nfrom tqdm.notebook import tqdm\nfrom joblib import Parallel, delayed\n\n\n#settings\npd.options.display.max_rows = 100\npd.options.display.max_columns = 100\n\n\nimport warnings\nwarnings.filterwarnings(\"ignore\")\n\nimport pytorch_lightning as pl\nrandom_seed=1234\npl.seed_everything(random_seed)","metadata":{"execution":{"iopub.status.busy":"2022-12-02T14:03:05.992232Z","iopub.execute_input":"2022-12-02T14:03:05.992758Z","iopub.status.idle":"2022-12-02T14:03:12.259146Z","shell.execute_reply.started":"2022-12-02T14:03:05.992705Z","shell.execute_reply":"2022-12-02T14:03:12.257799Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"

1. Data Description

\n\n

back to top

\n\n\n

[train/test]_images/[patient_id]/[image_id].dcm The mammograms, in dicom format. You can expect roughly 8,000 patients in the hidden test set. There are usually but not always 4 images per patient. Note that many of the images use the jpeg 2000 format which may you may need special libraries to load.

\nsample_submission.csv A valid sample submission. Only the first few rows are available for download.

\n[train/test].csv Metadata for each patient and image. Only the first few rows of the test set are available for download.

\n\n","metadata":{}},{"cell_type":"code","source":"train_df = pd.read_csv(\"/kaggle/input/rsna-breast-cancer-detection/train.csv\")\ntrain_df.head(3)","metadata":{"execution":{"iopub.status.busy":"2022-12-02T14:03:12.260641Z","iopub.execute_input":"2022-12-02T14:03:12.261384Z","iopub.status.idle":"2022-12-02T14:03:12.400297Z","shell.execute_reply.started":"2022-12-02T14:03:12.261349Z","shell.execute_reply":"2022-12-02T14:03:12.399128Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"

2. Load Original DICOM Images

\n

back to top

\n\n\n
    \n
  1. Cancer Positive Images
    2. \n
  2. Cancer Negative Images
    3. \n
  3. Unusually Difficult Cases
    4. \n
","metadata":{}},{"cell_type":"code","source":"def load_img(img_path, resize=True):\n #img_path=\"/kaggle/input/rsna-breast-cancer-detection/train_images/10038/1967300488.dcm\"\n dicom = pydicom.dcmread(img_path)\n img = dicom.pixel_array\n\n if resize:\n img = (img - img.min()) / (img.max() - img.min())\n\n if dicom.PhotometricInterpretation == \"MONOCHROME1\":\n img = 1 - img\n\n return img","metadata":{"execution":{"iopub.status.busy":"2022-12-02T14:03:12.40273Z","iopub.execute_input":"2022-12-02T14:03:12.403208Z","iopub.status.idle":"2022-12-02T14:03:12.410134Z","shell.execute_reply.started":"2022-12-02T14:03:12.40314Z","shell.execute_reply":"2022-12-02T14:03:12.408853Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def visualize_sample(\n base_path, \n data_df\n):\n nsamples = data_df.shape[0]\n ncols = 4\n nrows = int(np.ceil(nsamples/ncols))\n fig, axs = plt.subplots(nrows, ncols, figsize=(18, 6*nrows), sharey=False, sharex=False)\n #plt.style.use('seaborn-whitegrid')\n i = 0\n for _, row in data_df.iterrows():\n if axs.size<=ncols:\n ax = axs[i]\n else:\n ax = axs[i//ncols, i%ncols]\n patient_id = row['patient_id']\n image_id = row['image_id']\n img_path = f'{base_path}/{patient_id}/{image_id}.dcm'\n #data = load_dicom(img_path)\n data = load_img(img_path)\n ax.imshow(data)#, cmap=\"gray\"\n ax.set_title(f'patient: {patient_id}\\nimage:{image_id}\\nview:{row[\"view\"]}, side: {row[\"laterality\"]}\\ncancer positive: {row[\"cancer\"]}, age: {row[\"age\"]}', \n fontsize=13, loc='left')\n i = i + 1\n\n plt.suptitle(f\"\", fontsize=15,)\n plt.xticks(fontsize=10)\n plt.yticks(fontsize=10)\n plt.show()","metadata":{"_kg_hide-input":true,"_kg_hide-output":true,"execution":{"iopub.status.busy":"2022-12-02T14:03:12.412194Z","iopub.execute_input":"2022-12-02T14:03:12.412642Z","iopub.status.idle":"2022-12-02T14:03:12.424279Z","shell.execute_reply.started":"2022-12-02T14:03:12.412598Z","shell.execute_reply":"2022-12-02T14:03:12.423189Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"base_path=\"/kaggle/input/rsna-breast-cancer-detection/train_images\"","metadata":{"execution":{"iopub.status.busy":"2022-12-02T14:03:12.425536Z","iopub.execute_input":"2022-12-02T14:03:12.425829Z","iopub.status.idle":"2022-12-02T14:03:12.436206Z","shell.execute_reply.started":"2022-12-02T14:03:12.425803Z","shell.execute_reply":"2022-12-02T14:03:12.435214Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"

2.1 Cancer Positive

\n

back to top

","metadata":{}},{"cell_type":"code","source":"sub = train_df[(train_df['cancer']==1)].head(4)\n \nvisualize_sample(base_path=base_path, data_df=sub)","metadata":{"execution":{"iopub.status.busy":"2022-12-02T14:03:12.437572Z","iopub.execute_input":"2022-12-02T14:03:12.437971Z","iopub.status.idle":"2022-12-02T14:03:20.867253Z","shell.execute_reply.started":"2022-12-02T14:03:12.437942Z","shell.execute_reply":"2022-12-02T14:03:20.866027Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"sub","metadata":{"execution":{"iopub.status.busy":"2022-12-02T14:03:20.869134Z","iopub.execute_input":"2022-12-02T14:03:20.869995Z","iopub.status.idle":"2022-12-02T14:03:20.896242Z","shell.execute_reply.started":"2022-12-02T14:03:20.869947Z","shell.execute_reply":"2022-12-02T14:03:20.894503Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"

2.2 Cancer Negative

\n

back to top

","metadata":{}},{"cell_type":"code","source":"sub = train_df[(train_df['cancer']==0)].head(4)\n \nvisualize_sample(base_path=base_path, data_df=sub)","metadata":{"execution":{"iopub.status.busy":"2022-12-02T14:03:20.898161Z","iopub.execute_input":"2022-12-02T14:03:20.898591Z","iopub.status.idle":"2022-12-02T14:03:34.14434Z","shell.execute_reply.started":"2022-12-02T14:03:20.898555Z","shell.execute_reply":"2022-12-02T14:03:34.143243Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"

2.3 Unusually difficult cases

\n

back to top

","metadata":{}},{"cell_type":"code","source":"sub = train_df[(train_df['difficult_negative_case'])].head(4)\n \nvisualize_sample(base_path=base_path, data_df=sub)","metadata":{"execution":{"iopub.status.busy":"2022-12-02T14:03:34.149113Z","iopub.execute_input":"2022-12-02T14:03:34.149647Z","iopub.status.idle":"2022-12-02T14:03:39.23513Z","shell.execute_reply.started":"2022-12-02T14:03:34.1496Z","shell.execute_reply":"2022-12-02T14:03:39.233574Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"sub","metadata":{"execution":{"iopub.status.busy":"2022-12-02T14:03:39.236465Z","iopub.execute_input":"2022-12-02T14:03:39.236825Z","iopub.status.idle":"2022-12-02T14:03:39.254005Z","shell.execute_reply.started":"2022-12-02T14:03:39.236792Z","shell.execute_reply":"2022-12-02T14:03:39.253045Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"

3. Resize images using PIL package

\n

back to top

\n
\n\n
\n \n
\n\n","metadata":{}},{"cell_type":"code","source":"#Load PIL package\nimport PIL\nfrom PIL import Image\n","metadata":{"execution":{"iopub.status.busy":"2022-12-02T14:06:04.211096Z","iopub.execute_input":"2022-12-02T14:06:04.21158Z","iopub.status.idle":"2022-12-02T14:06:04.216908Z","shell.execute_reply.started":"2022-12-02T14:06:04.211547Z","shell.execute_reply":"2022-12-02T14:06:04.215653Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"img_path=\"/kaggle/input/rsna-breast-cancer-detection/train_images/10049/1207499426.dcm\"\ndicom = pydicom.dcmread(img_path)\nimg = dicom.pixel_array\n","metadata":{"execution":{"iopub.status.busy":"2022-12-02T14:03:39.255062Z","iopub.execute_input":"2022-12-02T14:03:39.255833Z","iopub.status.idle":"2022-12-02T14:03:39.703361Z","shell.execute_reply.started":"2022-12-02T14:03:39.255798Z","shell.execute_reply":"2022-12-02T14:03:39.702125Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"#https://github.com/pydicom/contrib-pydicom/blob/master/viewers/pydicom_PIL.py\ntry:\n import PIL.Image\n have_PIL = True\nexcept ImportError:\n have_PIL = False\n\ntry:\n import numpy as np\n have_numpy = True\nexcept ImportError:\n have_numpy = False\n\n\ndef get_LUT_value(data, window, level):\n \"\"\"Apply the RGB Look-Up Table for the given\n data and window/level value.\"\"\"\n if not have_numpy:\n raise ImportError(\"Numpy is not available.\"\n \"See http://numpy.scipy.org/\"\n \"to download and install\")\n\n return np.piecewise(data,\n [data <= (level - 0.5 - (window - 1) / 2),\n data > (level - 0.5 + (window - 1) / 2)],\n [0, 255, lambda data: ((data - (level - 0.5)) /\n (window - 1) + 0.5) * (255 - 0)])\n\n\ndef get_PIL_image(dataset):\n \"\"\"Get Image object from Python Imaging Library(PIL)\"\"\"\n if not have_PIL:\n raise ImportError(\"Python Imaging Library is not available. \"\n \"See http://www.pythonware.com/products/pil/ \"\n \"to download and install\")\n\n if ('PixelData' not in dataset):\n raise TypeError(\"Cannot show image -- DICOM dataset does not have \"\n \"pixel data\")\n # can only apply LUT if these window info exists\n if ('WindowWidth' not in dataset) or ('WindowCenter' not in dataset):\n bits = dataset.BitsAllocated\n samples = dataset.SamplesPerPixel\n if bits == 8 and samples == 1:\n mode = \"L\"\n elif bits == 8 and samples == 3:\n mode = \"RGB\"\n elif bits == 16:\n # not sure about this -- PIL source says is 'experimental'\n # and no documentation. Also, should bytes swap depending\n # on endian of file and system??\n mode = \"I;16\"\n else:\n raise TypeError(\"Don't know PIL mode for %d BitsAllocated \"\n \"and %d SamplesPerPixel\" % (bits, samples))\n\n # PIL size = (width, height)\n size = (dataset.Columns, dataset.Rows)\n\n # Recommended to specify all details\n # by http://www.pythonware.com/library/pil/handbook/image.htm\n im = PIL.Image.frombuffer(mode, size, dataset.PixelData,\n \"raw\", mode, 0, 1)\n\n else:\n ew = dataset['WindowWidth']\n ec = dataset['WindowCenter']\n ww = int(ew.value[0] if ew.VM > 1 else ew.value)\n wc = int(ec.value[0] if ec.VM > 1 else ec.value)\n image = get_LUT_value(dataset.pixel_array, ww, wc)\n # Convert mode to L since LUT has only 256 values:\n # http://www.pythonware.com/library/pil/handbook/image.htm\n im = PIL.Image.fromarray(image).convert('L')\n\n return im\n\n\ndef show_PIL(dataset):\n \"\"\"Display an image using the Python Imaging Library (PIL)\"\"\"\n im = get_PIL_image(dataset)\n im.show()","metadata":{"execution":{"iopub.status.busy":"2022-12-02T14:03:39.705468Z","iopub.execute_input":"2022-12-02T14:03:39.705818Z","iopub.status.idle":"2022-12-02T14:03:39.723128Z","shell.execute_reply.started":"2022-12-02T14:03:39.705786Z","shell.execute_reply":"2022-12-02T14:03:39.721752Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Step 1: convert numpy data to PIL image ","metadata":{}},{"cell_type":"code","source":"PIL_img=get_PIL_image(dicom)","metadata":{"execution":{"iopub.status.busy":"2022-12-02T14:03:39.724944Z","iopub.execute_input":"2022-12-02T14:03:39.725363Z","iopub.status.idle":"2022-12-02T14:03:39.87072Z","shell.execute_reply.started":"2022-12-02T14:03:39.725313Z","shell.execute_reply":"2022-12-02T14:03:39.869506Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Step 2: show PIL image size (as well as original image size)","metadata":{}},{"cell_type":"code","source":"print(f\"PIL image size: {PIL_img.size}, PIL image height: {PIL_img.height}, \\\nPIL image width: {PIL_img.width}, Original image shape: {img.shape}\")","metadata":{"execution":{"iopub.status.busy":"2022-12-02T14:03:39.872204Z","iopub.execute_input":"2022-12-02T14:03:39.872555Z","iopub.status.idle":"2022-12-02T14:03:39.879081Z","shell.execute_reply.started":"2022-12-02T14:03:39.872524Z","shell.execute_reply":"2022-12-02T14:03:39.877545Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Step 3: compare the original and PIL images","metadata":{}},{"cell_type":"code","source":"fig, axes = plt.subplots(nrows=1, ncols=2, figsize=(14, 8))\n\n\naxes[0].imshow(img)\naxes[0].set_title(f'Original Image')\naxes[1].imshow(PIL_img)\naxes[1].set_title(f'PIL Image')\n\nplt.suptitle(f\"compare original image and PIL image\", fontsize=15,)\nplt.xticks(fontsize=10)\nplt.yticks(fontsize=10)\nplt.show()","metadata":{"execution":{"iopub.status.busy":"2022-12-02T14:03:39.880991Z","iopub.execute_input":"2022-12-02T14:03:39.881451Z","iopub.status.idle":"2022-12-02T14:03:42.061892Z","shell.execute_reply.started":"2022-12-02T14:03:39.881409Z","shell.execute_reply":"2022-12-02T14:03:42.060614Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Step 4: resize image using PIL thumbnail function","metadata":{}},{"cell_type":"code","source":"#create the thumbnail of the image\n\nif hasattr(Image, 'Resampling'): # Pillow<8.4.0\n PIL_img.thumbnail((1024, 1024), resample=Image.Resampling.LANCZOS, reducing_gap=10)\n if (PIL_img.height< PIL_img.width):\n PIL_img = PIL_img.transpose(PIL.Image.Transpose.ROTATE_90)\nelse:\n PIL_img.thumbnail((1024, 1024), resample=Image.LANCZOS, reducing_gap=10)\n if (PIL_img.height> img.width):\n PIL_img = PIL_img.transpose(PIL.Image.ROTATE_90)\n","metadata":{"execution":{"iopub.status.busy":"2022-12-02T14:06:09.677986Z","iopub.execute_input":"2022-12-02T14:06:09.678454Z","iopub.status.idle":"2022-12-02T14:06:09.770411Z","shell.execute_reply.started":"2022-12-02T14:06:09.678414Z","shell.execute_reply":"2022-12-02T14:06:09.769197Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"Step 5: compare the original and PIL resized images","metadata":{}},{"cell_type":"code","source":"print(f\"PIL image size: {PIL_img.size}, PIL image height: {PIL_img.height}, \\\nPIL image width: {PIL_img.width}, Original image shape: {img.shape}\")\n\nfig, axes = plt.subplots(nrows=1, ncols=2, figsize=(14, 8))\n\n\naxes[0].imshow(img)\naxes[0].set_title(f'Original Image')\naxes[1].imshow(PIL_img)\naxes[1].set_title(f'PIL resized Image')\n\nplt.suptitle(f\"compare original image and PIL image\", fontsize=15,)\nplt.xticks(fontsize=10)\nplt.yticks(fontsize=10)\nplt.show()","metadata":{"execution":{"iopub.status.busy":"2022-12-02T14:06:11.590655Z","iopub.execute_input":"2022-12-02T14:06:11.591081Z","iopub.status.idle":"2022-12-02T14:06:13.0096Z","shell.execute_reply.started":"2022-12-02T14:06:11.591047Z","shell.execute_reply":"2022-12-02T14:06:13.008359Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"\n\n

4. Image Augmentation by torchvisaion tranforms

\n

back to top

\n\n
    \n
  1. 4.1. center crop the image and pad image
    2. \n
  2. 4.2. Add Gaussian Blur to images
    3. \n
\n
\n\n
\n \n
","metadata":{}},{"cell_type":"code","source":"import torch\nimport torch.nn as nn\nimport torch.optim as optim\nfrom torch.optim import lr_scheduler\nimport torch.backends.cudnn as cudnn\nimport numpy as np\nimport torchvision\nfrom torchvision import datasets, models, transforms\nimport matplotlib.pyplot as plt","metadata":{"execution":{"iopub.status.busy":"2022-12-02T14:06:15.917514Z","iopub.execute_input":"2022-12-02T14:06:15.918417Z","iopub.status.idle":"2022-12-02T14:06:15.927249Z","shell.execute_reply.started":"2022-12-02T14:06:15.918326Z","shell.execute_reply":"2022-12-02T14:06:15.925349Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"

4.1. center crop the image and pad image

\n

back to top

","metadata":{}},{"cell_type":"code","source":"#https://stackoverflow.com/questions/10965417/how-to-convert-a-numpy-array-to-pil-image-applying-matplotlib-colormap\n\n#use torchvision to center crop the image\nimg2 = transforms.functional.center_crop(PIL_img, 1024)\n#use torchvision to pad the image\nimg3 = transforms.functional.pad(PIL_img, 10)","metadata":{"execution":{"iopub.status.busy":"2022-12-02T14:06:16.57954Z","iopub.execute_input":"2022-12-02T14:06:16.580019Z","iopub.status.idle":"2022-12-02T14:06:16.590025Z","shell.execute_reply.started":"2022-12-02T14:06:16.579983Z","shell.execute_reply":"2022-12-02T14:06:16.587854Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"print(f\"PIL image size: {PIL_img.size}, PIL image height: {PIL_img.height},\\n \\\nPIL image width: {PIL_img.width}, Original image shape: {img.shape},\\n \\\nCenter crop image shape: {img2.size}, Pad image shape: {img3.size},\\n \\\n\")\n\nfig, axes = plt.subplots(nrows=1, ncols=4, figsize=(18, 8))\n\n\naxes[0].imshow(img)\naxes[0].set_title(f'Original Image')\naxes[1].imshow(PIL_img)\naxes[1].set_title(f'PIL resized Image')\naxes[2].imshow(img2)\naxes[2].set_title(f'Center crop Image')\naxes[3].imshow(img3)\naxes[3].set_title(f'Pad Image')\n\nplt.suptitle(f\"compare original image, PIL image, center crop and pad images\", fontsize=15,)\nplt.xticks(fontsize=10)\nplt.yticks(fontsize=10)\nplt.show()\n","metadata":{"execution":{"iopub.status.busy":"2022-12-02T14:06:17.338244Z","iopub.execute_input":"2022-12-02T14:06:17.338714Z","iopub.status.idle":"2022-12-02T14:06:19.091508Z","shell.execute_reply.started":"2022-12-02T14:06:17.338679Z","shell.execute_reply":"2022-12-02T14:06:19.08966Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"\n

4.2. Add Gaussian Blur to images

\n

back to top

","metadata":{}},{"cell_type":"code","source":"#https://dsp.stackexchange.com/questions/10057/gaussian-blur-standard-deviation-radius-and-kernel-size\n#try different kernal sizes\n\nfig, axes = plt.subplots(nrows=1, ncols=4, figsize=(18, 8))\n\naxes[0].imshow(PIL_img)\naxes[0].set_title(f'PIL image')\n\nfor i, k_sizes in enumerate([(5, 5),(15, 15), (25, 45)]):\n img4 = transforms.functional.gaussian_blur(PIL_img, kernel_size=k_sizes, sigma=(0.1, 5))\n #print(i+1, k_sizes, img4.size, np.asarray(img4, np.uint8).min(), np.asarray(img4, np.uint8).max())\n axes[i+1].imshow(img4)\n axes[i+1].set_title(f'Kernal sizes: {k_sizes}')\n\nplt.suptitle(f\"compare PIL image and different Gaussian Blur Kernel sizes\", fontsize=15,)\nplt.xticks(fontsize=10)\nplt.yticks(fontsize=10)\nplt.show()\n","metadata":{"execution":{"iopub.status.busy":"2022-12-02T14:06:19.094154Z","iopub.execute_input":"2022-12-02T14:06:19.095358Z","iopub.status.idle":"2022-12-02T14:06:20.770551Z","shell.execute_reply.started":"2022-12-02T14:06:19.095292Z","shell.execute_reply":"2022-12-02T14:06:20.769219Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"#try different sigma\n\nfig, axes = plt.subplots(nrows=1, ncols=4, figsize=(18, 8))\n\naxes[0].imshow(PIL_img)\naxes[0].set_title(f'PIL image')\n\nfor i, sigma in enumerate([ (0.05, 1), (0.1, 5),(0.8, 10)]):\n img4 = transforms.functional.gaussian_blur(PIL_img, kernel_size=(5, 9), sigma=sigma)\n #print(i, k_sizes, img4.size, np.asarray(img4, np.uint8).min(), np.asarray(img4, np.uint8).max())\n axes[i+1].imshow(img4)\n axes[i+1].set_title(f'sigma: {sigma}')\n\nplt.suptitle(f\"compare PIL image and different Gaussian Blur sigma\", fontsize=15,)\nplt.xticks(fontsize=10)\nplt.yticks(fontsize=10)\nplt.show()\n","metadata":{"execution":{"iopub.status.busy":"2022-12-02T14:06:20.772596Z","iopub.execute_input":"2022-12-02T14:06:20.773086Z","iopub.status.idle":"2022-12-02T14:06:21.774833Z","shell.execute_reply.started":"2022-12-02T14:06:20.773046Z","shell.execute_reply":"2022-12-02T14:06:21.773967Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"#try different sigma and kernel_size\n\nfig, axes = plt.subplots(nrows=1, ncols=4, figsize=(18, 8))\n\naxes[0].imshow(PIL_img)\naxes[0].set_title(f'PIL image')\n\nfor i, (sigma, k_sizes) in enumerate([[(0.1, 5), (35, 65)],\n [(0.8, 10), (5, 9)], [(0.8, 10), (35, 65)]]):\n img4 = transforms.functional.gaussian_blur(PIL_img, kernel_size=k_sizes, sigma=sigma)\n #print(i, k_sizes, img4.size, np.asarray(img4, np.uint8).min(), np.asarray(img4, np.uint8).max())\n axes[i+1].imshow(img4)\n axes[i+1].set_title(f'kernel_size={k_sizes}, sigma={sigma}')\n\nplt.suptitle(f\"compare PIL image and different Gaussian Blur kernel size and sigma\", fontsize=15,)\nplt.xticks(fontsize=10)\nplt.yticks(fontsize=10)\nplt.show()\n","metadata":{"execution":{"iopub.status.busy":"2022-12-02T14:06:21.77624Z","iopub.execute_input":"2022-12-02T14:06:21.776787Z","iopub.status.idle":"2022-12-02T14:06:25.699821Z","shell.execute_reply.started":"2022-12-02T14:06:21.776753Z","shell.execute_reply":"2022-12-02T14:06:25.698141Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"\n

Reference:

\n\n* https://www.kaggle.com/code/ihelon/brain-tumor-eda-with-animations-and-modeling\n* https://www.kaggle.com/code/theoviel/dicom-resized-png-jpg/notebook\n* https://www.kaggle.com/code/xxxxyyyy80008/strip-ai-understand-image-processing","metadata":{}}]}