Use Case: medical image segmentation

[lorenzoh]

This use case covers this MonAI tutorial for 3D multi-class semantic segmentation.

Below is a list of references to parts that can be done and functionality that is missing to recreate the tutorial in Julia.

Dataset/data container

Available from http://medicaldecathlon.com/

ToDos:

• create a data container (see https://lorenzoh.github.io/DataLoaders.jl/dev/docs/datacontainers.html)
• one observation is one 3D image and multiple 3D masks
• select orientation 3d image is read in

Data preprocessing/transformation/augmentation

• crop/resize/scale
• normalization
• random jitter

See DataAugmentation.jl

ToDos:

• item implementation for 3d iamges and masks
• affine transformations support 3D images and 3D masks
• random jitter transform

Data loading/data iterator

Use DataLoaders.jl

For wrapping the whole data pipeline, consider DLPipelines.jl

Model

see https://github.com/DhairyaLGandhi/UNet.jl, should probably work fine. Maybe add some tweaks from fastai's implementation.

Training

• dice loss function
• optimizer: use Flux.ADAM
• training loop: use FluxTraining.jl
• for logs and visualizations, use Logger with TensorBoardBackend
• use Checkpointer to save the model

ToDos

• implement dice in FluxTraining.jl

[darsnack]

I am closing this after moving it to a discussion topic. The idea is to reduce the issues section to actionable items for contributors. Discussions and user stories can go in the Discussions tab.