create_patches.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "source": [
    "This script will take an array of 4k x 4k TMA spots as input and produce patches\r\n",
    "for training of the MIL model. Each instance will inherit the label of its bag.\r\n",
    "\r\n",
    "Assumptions for this script to work:\r\n",
    "* the following file structure:\r\n",
    "\r\n",
    "\r\n",
    "ROOT/\r\n",
    "\r\n",
    "\t├── tma_spots\r\n",
    "\r\n",
    "    \t\t├── slide_1.png\r\n",
    "\r\n",
    "    \t\t├── slide_2.png\r\n",
    "\r\n",
    "    \t\t└── ...\r\n",
    "\t\r\n",
    "\t└── data\r\n",
    "\t\t\t└──{dataname}_bag_labels.csv\r\n",
    "\r\n",
    "\t\t\t\r\n",
    "Output:\r\n",
    "\r\n",
    "{dataname}_train.pytable\r\n",
    "\r\n",
    "{dataname}_val.pytable\r\n",
    "\r\n",
    "in data directory\r\n",
    "\r\n",
    "TODO:\r\n",
    "* Save slide-level id and label for each of the input images (TMA spots)"
   ],
   "metadata": {}
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "source": [
    "# --- parameters\r\n",
    "dataname = 'tma_patches'\r\n",
    "patch_size = 256\r\n",
    "stride_size = 256\r\n",
    "tma_spot_directory = './tma_spots'\r\n",
    "test_set_size = 0.1\r\n",
    "classes = [0,1]\r\n",
    "mirror_pad_size = 128\r\n",
    "resize = 1"
   ],
   "outputs": [],
   "metadata": {}
  },
  {
   "cell_type": "markdown",
   "source": [
    "IMPORTS"
   ],
   "metadata": {}
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "source": [
    "import torch\r\n",
    "import tables\r\n",
    "\r\n",
    "import os,sys\r\n",
    "import glob\r\n",
    "\r\n",
    "import PIL\r\n",
    "import numpy as np\r\n",
    "\r\n",
    "import cv2\r\n",
    "import matplotlib.pyplot as plt\r\n",
    "import csv\r\n",
    "\r\n",
    "from sklearn import model_selection\r\n",
    "import sklearn.feature_extraction.image\r\n",
    "import random\r\n",
    "\r\n",
    "seed = random.randrange(sys.maxsize) #get a random seed so that we can reproducibly do the cross validation setup\r\n",
    "random.seed(seed) # set the seed\r\n",
    "print(f\"random seed (note down for reproducibility): {seed}\")"
   ],
   "outputs": [],
   "metadata": {}
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "source": [
    "img_dtype = tables.UInt8Atom()  # dtype in which the images will be saved, this indicates that images will be saved as unsigned int 8 bit, i.e., [0,255]\r\n",
    "filenameAtom = tables.StringAtom(itemsize=255)\r\n",
    "\r\n",
    "#get the raw TMA spot files\r\n",
    "files = glob.glob(file_directory)\r\n",
    "\r\n",
    "# convert bag_labels csv file into a list of file names and associated labels.\r\n",
    "# assumes bag_labels.csv has two columns (slide_id and label). Fname should contain the following set {0, 1, ..., nSPOTS}.\r\n",
    "tmp = open('./data/{dataname}_bag_labels.csv', 'r')\r\n",
    "bag_labels = csv.DictReader(tmp)\r\n",
    "bag_labels = list(bag_labels)\r\n",
    "\r\n",
    "#create training and validation stages and split the files appropriately between them\r\n",
    "phases={}\r\n",
    "phases[\"train\"],phases[\"val\"]=next(iter(model_selection.ShuffleSplit(n_splits=1,test_size=test_set_size).split(files)))"
   ],
   "outputs": [],
   "metadata": {}
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "source": [
    "storage={} #holder for future pytables\r\n",
    "\r\n",
    "block_shape={} #block shape specifies what we'll be saving into the pytable array, here we assume that masks are 1d and images are 3d\r\n",
    "block_shape[\"img\"]= np.array((patch_size,patch_size,3))\r\n",
    "\r\n",
    "filters=tables.Filters(complevel=6, complib='zlib') #we can also specify filters, such as compression, to improve storage speed\r\n",
    "\r\n",
    "for phase in phases.keys(): #now for each of the phases, we'll loop through the files\r\n",
    "\tprint(phase)\r\n",
    "\r\n",
    "\ttotals = np.zeros((2,len(classes)))\r\n",
    "\ttotals[0,:] = classes\r\n",
    "\r\n",
    "\thdf5_file = tables.open_file(f\".data/{dataname}_{phase}.pytable\", mode='w') #open the respective pytable\r\n",
    "\tstorage[\"slide_ids\"] = hdf5_file.create_earray(hdf5_file.root, 'slide_ids', filenameAtom, (0,)) #create the array for storage\r\n",
    "\r\n",
    "\tstorage['img'] = hdf5_file.create_earray(hdf5_file.root, 'img', img_dtype,  \r\n",
    "                                            shape=np.append([0],block_shape['img']), \r\n",
    "                                            chunkshape=np.append([1],block_shape['img']),\r\n",
    "                                            filters=filters)\r\n",
    "\tstorage[\"labels\"]= hdf5_file.create_earray(hdf5_file.root, \"labels\", img_dtype,  \r\n",
    "\t\t\t\t\t\t\t\t\t\t\tshape=[0], \r\n",
    "\t\t\t\t\t\t\t\t\t\t\tchunkshape=[1],\r\n",
    "\t\t\t\t\t\t\t\t\t\t\tfilters=filters)\r\n",
    "\r\n",
    "\tfor filei in phases[phase]:\r\n",
    "\t\tfname = files[filei]\r\n",
    "\r\n",
    "\t\tprint(fname)\r\n",
    "\t\tio=cv2.cvtColor(cv2.imread(\"./imgs/\"+os.path.basename(fname).replace(\"_mask.png\",\".tif\")),cv2.COLOR_BGR2RGB)\r\n",
    "\t\tinterp_method=PIL.Image.BICUBIC\r\n",
    "\r\n",
    "\t\tio = cv2.resize(io,(0,0),fx=resize,fy=resize, interpolation=interp_method) #resize it as specified above\r\n",
    "\t\tio = np.pad(io, [(mirror_pad_size, mirror_pad_size), (mirror_pad_size, mirror_pad_size), (0, 0)], mode=\"reflect\")\r\n",
    "\t\t\r\n",
    "\t\t#convert input image into overlapping tiles, size is ntiler x ntilec x 1 x patch_size x patch_size x3\r\n",
    "\t\tio_arr_out=sklearn.feature_extraction.image.extract_patches(io,(patch_size,patch_size,3),stride_size)\r\n",
    "\r\n",
    "\t\t#resize it into a ntile x patch_size x patch_size x 3\r\n",
    "\t\tio_arr_out=io_arr_out.reshape(-1,patch_size,patch_size,3)\r\n",
    "\r\n",
    "\t\t#save the 4d tensor to a table\r\n",
    "\t\tstorage['img'].append(io_arr_out)\r\n",
    "\r\n",
    "\t\tstorage['slide_ids'].append([fname for x in range(io_arr_out.shape[0])])\r\n",
    "\r\n",
    "\t\tstorage['label'].append(bag_labels[fname]['label'])\r\n",
    "\t\t\r\n",
    "\t\t#script to check whether fname matches the entry in the fname column\r\n",
    "\t\tprint(fname + '\\n')\r\n",
    "\t\tprint(bag_labels[fname]['slide_id'])\r\n",
    "\r\n",
    "\t#lastly, we should store the number of pixels\r\n",
    "\tnpixels=hdf5_file.create_carray(hdf5_file.root, 'numpixels', tables.Atom.from_dtype(totals.dtype), totals.shape)\r\n",
    "\tnpixels[:]=totals\r\n",
    "\thdf5_file.close()\r\n"
   ],
   "outputs": [],
   "metadata": {}
  }
 ],
 "metadata": {
  "orig_nbformat": 4,
  "language_info": {
   "name": "python"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}