3D U-Net in TensorFlow

Author: Daniel Homola

Main deliverables:

• Report
• Data exploration notebook
• Model exploration notebook

Overview

MRI scans from 70 patients were used to learn to automatically segment the 3D volume of scans, and therefore spatially identify the outlines of the central gland (CG) and peripheral zone (PZ).

The aim of this project was to quickly establish a benchmark model, with minimal / lightweight code, only relying on core TensorFlow and Python, i.e. without using Keras or other wrapper libraries.

Dataset used

• NCI-ISBI 2013 Challenge: Automated Segmentation of Prostate Structures

Original Paper

• 3D U-Net: Learning Dense Volumetric Segmentation from Sparse Annotation

Data

Prostate MRI scans with segmentation

Objective: assign mutually exclusive class labels to each pixel/voxel.

Class labels: 0: background, 1: central gland , 2: peripheral zone

Input	Shape	Explanation	Example
X: 5-D Tensor	(?, ?, 128, 128, 1)	Resized (128, 128) MRI scans with depth (between 15 and 40) and a single channel. Variable batch size.
y: 5-D Tensor	(?, ?, 128, 128, 3)	Resized (128, 128) segmentation images with depth corresponding to X and a three classes, i.e. 3 channels one-hot encoded. Variable batch size.

Get Started

Setup project

• Create new Python3 virtualenv (assumes you have virtualenv and virtualenvwrapper installed and set up)
• Install dependencies.

mkvirtualenv --python=`which python3` unet3d
workon unet3d
make requirements

Additional helper functions can be explored with.

make help

Download dataset

• The dataset was downloaded from here.
• As per the instructions, the training (60) and leaderboard (10) subjects were pooled to form the train dataset (1816 scans in total).
• The test dataset consists of 10 patients with 271 scans.
• Unzip them and place them in data/raw.

Explore and preprocess data

• In this notebook, we load in the MRI scans and their segmentations, build a Dataset object for the train and test set.
• Then we check some basic stats of the datasets and visualise a few scans.
• Finally, we carry out our preprocessing steps and save the train and test datasets.

jupyter notebook "notebooks/data_exploration.ipynb"

Build model

Train and evaluate

Train model on train set and evaluate it on test set using the base model architecture.

python src/main.py -model_dir models/base_model -mode train_eval

Do the same with same parameters but without batch normalisation.

python src/main.py -model_dir models/base_model_no_bn -mode train_eval

For more options:

python src/main.py -h

usage: main.py [-h] -model_dir MODEL_DIR [-mode MODE]
               [-pred_ix PRED_IX [PRED_IX ...]]

Train, eval, predict 3D U-Net model.

optional arguments:
  -h, --help            show this help message and exit
  -model_dir MODEL_DIR  Experiment directory containing params.json
  -mode MODE            One of train, train_eval, eval, predict.
  -pred_ix PRED_IX [PRED_IX ...]
                        Space separated list of indices of patients to
                        predict.
usage: main.py [-h] -model_dir MODEL_DIR [-mode MODE]
               [-pred_ix PRED_IX [PRED_IX ...]]

Train, eval, predict 3D U-Net model.

optional arguments:
  -h, --help            show this help message and exit
  -model_dir MODEL_DIR  Experiment directory containing params.json
  -mode MODE            One of train, train_eval, eval, predict.
  -pred_ix PRED_IX [PRED_IX ...]
                        Space separated list of indices of patients to
                        predict.

Predict

Predict all 10 patients in test set with the trained base model and save their predictions to model directory.

python src/main.py -model_dir models/base_model -mode predict - pred_ix 0 1 2 3 4 5 6 7 8 9

Explore predictions

• Open the Jupyter notebook to have a look at test cases

jupyter notebook "notebooks/model_exploration.ipynb"

Example predictions with the base model

Ground truth is on left, prediction is on the right in each image.

Example predictions with the deeper model

Ground truth is on left, prediction is on the right in each image.