{ "cells": [ { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "%matplotlib inline" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "\npynet metastasis tumor segmentation\n===================================\n\nCredit: A Grigis\n\npynet is a Python package related to deep learning and its application in\nMRI mediacal data analysis. It is accessible to everybody, and is reusable\nin various contexts. The project is hosted on github:\nhttps://github.com/neurospin/pynet.\n\nIn this notebook we will learn how to segment tumors using MRI images from the\nBrats dataset. The NvNet proposed by Andriy Myronenko's will be trained. This\nnetwork is a combination of a Vnet (3d Unet) and a VAE (variation\nauto-encoder).\n\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "import os\nimport sys\nif \"CI_MODE\" in os.environ:\n sys.exit()" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "Inspect the NvNet network\n--------------------------\n\nInspect some layers of the network.\n\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "from pprint import pprint\nimport pynet.models as models\nfrom pynet import NetParameters\nfrom pynet.utils import get_named_layers\nfrom pynet.utils import setup_logging\n# from pynet.plotting.network import plot_net_rescue\n\nsetup_logging(level=\"info\")\n\nmodel = models.NvNet(\n input_shape=(128, 128, 128),\n in_channels=1,\n num_classes=4,\n activation=\"relu\",\n normalization=\"group_normalization\",\n mode=\"trilinear\",\n with_vae=True)\nlayers = get_named_layers(model)\npprint(layers)\n# graph_file = plot_net_rescue(model, (1, 1, 128, 128, 128))" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "Import the brats dataset\n------------------------\n\nUse the fetcher of the pynet package.\n\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "from pynet.datasets import DataManager, fetch_brats\nfrom pynet.plotting import plot_data\n\ndata = fetch_brats(\n datasetdir=\"/neurospin/nsap/datasets/brats\")\nmanager = DataManager(\n input_path=data.input_path,\n metadata_path=data.metadata_path,\n output_path=data.output_path,\n projection_labels=None,\n number_of_folds=10,\n batch_size=1,\n stratify_label=\"grade\",\n # input_transforms=[\n # RandomFlipDimensions(ndims=3, proba=0.5, with_channels=True),\n # Offset(nb_channels=4, factor=0.1)],\n sampler=\"random\",\n add_input=True,\n test_size=0.1,\n pin_memory=True)\ndataset = manager[\"test\"][:1]\nprint(dataset.inputs.shape, dataset.outputs.shape)\nplot_data(dataset.inputs, channel=1, nb_samples=5)\nplot_data(dataset.outputs, channel=1, nb_samples=5)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "Training\n--------\n\nFrom the available models load the 3D NvNet, and start the training.\n\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "import os\nfrom torch.optim import lr_scheduler\nimport pynet\nfrom pynet.losses import NvNetCombinedLoss\nfrom pynet.interfaces import NvNetSegmenter\nfrom pynet.plotting import plot_history\nfrom pynet.history import History\n\nlosses = pynet.get_tools(tool_name=\"losses\")\nmy_loss = losses[\"NvNetCombinedLoss\"](\n num_classes=4,\n k1=0.1,\n k2=0.1)\noutdir = \"/neurospin/nsap/tmp/nvnet\"\nif not os.path.isdir(outdir):\n os.mkdir(outdir)\ntrained_model = os.path.join(outdir, \"model_0_epoch_99.pth\")\nnvnet_params = NetParameters(\n input_shape=(150, 190, 135),\n in_channels=4,\n num_classes=4,\n activation=\"relu\",\n normalization=\"group_normalization\",\n mode=\"trilinear\",\n with_vae=True)\n\nif os.path.isfile(trained_model):\n nvnet = NvNetSegmenter(\n nvnet_params,\n optimizer_name=\"Adam\",\n learning_rate=1e-4,\n weight_decay=1e-5,\n loss=my_loss,\n pretrained=trained_model,\n use_cuda=True)\n train_history = History.load(\n os.path.join(outdir, \"train_0_epoch_9.pkl\"))\n valid_history = History.load(\n os.path.join(outdir, \"validation_0_epoch_9.pkl\"))\nelse:\n nvnet = NvNetSegmenter(\n nvnet_params,\n optimizer_name=\"Adam\",\n learning_rate=1e-4,\n weight_decay=1e-5,\n loss=my_loss,\n use_cuda=True)\n scheduler = lr_scheduler.ReduceLROnPlateau(\n optimizer=nvnet.optimizer,\n mode=\"min\",\n factor=0.5,\n patience=5)\n train_history, valid_history = nvnet.training(\n manager=manager,\n nb_epochs=100,\n checkpointdir=outdir,\n # fold_index=0,\n scheduler=scheduler,\n with_validation=True)\nprint(train_history)\nprint(valid_history)\nplot_history(train_history)" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "Testing\n-------\n\nFinaly use the testing set and check the results.\n\n" ] }, { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "y_pred, X, y_true, loss, values = nvnet.testing(\n manager=manager,\n with_logit=False,\n predict=False)\nprint(y_pred.shape, X.shape, y_true.shape)\n# y_pred = np.expand_dims(y_pred, axis=1)\n# data = np.concatenate((y_pred, y_true, X), axis=1)\n# plot_data(data, nb_samples=5)\n\nif \"CI_MODE\" not in os.environ:\n import matplotlib.pyplot as plt\n plt.show()" ] } ], "metadata": { "kernelspec": { "display_name": "Python 3", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.6.12" } }, "nbformat": 4, "nbformat_minor": 0 }