MICCAI 2022 Tutorial

Weakly Supervised CNN Segmentation: Models and Optimization

The code for the 2021 tutorial is available at commit c8d5d248dd8b504b53c56449ea526799df87e3ea (tag MICCAI2021). Other older tutorials (MICCAI2020, MICCAI2019) are also tagged.

This repository contains the code of the hand-on tutorial. The hands-on will be done in three main parts:

• naive sizeloss, to introduce the pipeline and general methodology;
• combined size and centroid supervision, with a quadratic penalty;
• combined size and centroid supervision, with an extended log-barrier.

Hands-on

The first goal is to enforce some inequality constraints on the size of the predicted segmentation in the form:

lower bound <= predicted size <= upper bound

where predicted size is the sum of all predicted probabilities (softmax) over the whole image.

To make the example simpler, we will define the lower and upper bounds to 0.9 and 1.1 times the ground truth size. All the code is contained within the code/ folder. The following assume you moved in that directory.

Requirements

The code has those following dependencies:

python3.7+
pytorch (latest)
torchvision
numpy
tqdm

Running the PROMISE12 example requires some additional packages:

simpleitk
scikit-image
PIL

ACDC relies on, for the slicing:

nibabel

To quickly bootstrap an installation, you can use the following commands in your shell:

python3 -m venv miccaiweaksup
source miccaiweaksup/bin/activate
which python  # ensure this is not your system's python anymore
python3 -m pip install -r requirements.txt

and at the end of the session, in your shell:

deactivate

Data

The data for the toy example is stored in code/data/TOY. If you wish, you can regenerate the dataset with:

make -B data/TOY
make -B data/TOY2

or you can use gen_toy.py directly.

Participants willing to try the PROMISE12 setting need to download the data themselves, then put the .zip inside the code/data folder (a list of files is available in code/data/promise12.lineage). Once the three files are there, the slicing into 2D png files is automated:

make data/PROMISE12

It will:

• checks data integrity
• extract the zip
• slice into 2d slices
• generate weak labels from the actual ground truth

The same goes for ACDC:

make data/ACDC

Training

>>> ./main.py -h
usage: main.py [-h] [--epochs EPOCHS] [--dataset {TOY,TOY2,PROMISE12,ACDC}] [--mode {constrained,unconstrained,full}] [--gpu]

options:
  -h, --help            show this help message and exit
  --epochs EPOCHS
  --dataset {TOY,TOY2,PROMISE12,ACDC}
  --mode {constrained,unconstrained,full}
  --gpu

The toy example is designed to run under 5 minutes on a laptop, training on CPU. The following commands are equivalent

python3 main.py
./main.py
./main.py --epochs 200 --dataset TOY --mode unconstrained

The three modes correspond to:

• unconstrained: use the weak labels, with only a partial cross-entropy (won't learn anything)
• constrained: use the weak labels, with partial cross-entropy + size constraint (will learn)
• full: use full labels, with cross entropy (will learn, for obvious reasons)

The settings for PROMISE12 are too simple to get state of the art results, even in the full mode, but it gives a good starting point for new practitioners to then build on.

Examples constraining both the size and the centroid, without resorting to any pixel-wise supervision, are shown:

>>>  ./main_centroid.py -h
usage: main_centroid.py [-h] [--epochs EPOCHS] [--dataset {TOY2}] [--mode {quadratic,logbarrier}] [--gpu]

options:
  -h, --help            show this help message and exit
  --epochs EPOCHS
  --dataset {TOY2}
  --mode {quadratic,logbarrier}
  --gpu

Going further

The current code can have a lot of assert, which is very helpful for development and debugging, but might slow the training a bit. assert can be disabled completely, by adding the, -O parameter to the python interpreter:

python3 -O main.py --dataset PROMISE12 --mode full --gpu

You can always e-mail me for technical questions.

Useful resources

• Custom segmentation viewer
• Pytorch documentation
• Numpy documentation
• GNU Make manual

MICCAI 2021 recordings

While the recording and slides are not yet available, the ones from last year are still online.

Slides

Slides from the three sessions are available in the slides/ folder.

Recordings

• Session 1: Structure-driven priors: Regularization
• Session 2: Knowledge-driven priors (e.g., anatomy): Constraints
• Session 3: Data-driven priors: Adversarial learning
• Session 4: Hands-on: Size constraints