PUMA Challenge Baseline Track 1

This repository contains a Dockerized environment for running the PUMA Challenge Evaluation track using CUDA 12.1. The container includes all necessary dependencies to execute the model and run inference on input data.

Prerequisites

• Docker (Ensure that Docker is installed and supports GPU with CUDA 12.1 or newer)
• NVIDIA Docker Toolkit for GPU support

Build the container

You can build the Docker image using the build.sh script. Ensure GPU support is enabled.

Adding the weights

Weights can be downloaded from: https://zenodo.org/records/13881999

• The content of Hover-NeXt_3_classes needs to be placed in the checkpoint folder.
• The nnU-net/checkpoint_best.pth needs to be placed inside the folder: \nnunetv2\nnunetv2_hist\nnUNet_results\Dataset526_Mark\nnUNetTrainer_nnUNetPlans_2d\fold_4.

Running the container.

Use the test_run.sh script to run the container.

Input & Output

One input file will be mounted per container at (algorithm job) /input/images/melanoma-whole-slide-image/<uuid>.tif.

Two output files are expected files inside the /output directory:

• melanoma-3-class-nuclei-segmentation.json contains the nuclei predictions in "Multiple Polygons" format.
• images/melanoma-tissue-mask-segmentation/<uuid>.tif contains the tissue predictions, where pixels should be given the following values: 'Background': 0, 'Stroma': 1, 'Blood Vessel': 2, 'Tumor': 3, 'Epidermis': 4, and 'Necrosis': 5

In the /test directory, one example input case can be found.

Saving the container

Use save.sh to save the container.