Objective

The primary objective of this challenge is to develop a highly accurate deep learning model capable of segmenting the Aorta, its branches, and associated zones from Computed Tomography Angiography (CTA) volumes. Participants will work with medical imaging data to create robust algorithms that can significantly assist in clinical diagnosis and treatment planning.

Model Development Phase

• Dataset Access: Starting from April 15, 2024, participants will receive access to the dataset upon completion and submission of the data-sharing agreement form.
• Dataset Characteristics:
• The dataset comprises 50 CTA images with the following specifications
• Axial size: Ranging from a minimum of 389x389 pixels to a maximum of 516x516 pixels, with an average of 450x450 pixels.
• Isotropic voxel resolution: Uniform at (1mm, 1mm, 1mm).
• Number of axial slices: Varies between 578 to 801, with an average of 695 slices.
• Inspiration from GitHub Repository: Participants are encouraged to draw inspiration from our GitHub repository, which provides an example utilizing the state-of-the-art SwinUNETR architecture for training purposes.
• Evaluation on Unseen Data: Starting from May 20, 2024, participants can submit their Docker container image to evaluate it on 10 unseen CTA images. Hints for evaluation can be obtained from our provided evaluation code.
• Algorithm Refinement Period: From May 20 to July 14, 2024, participants can refine their algorithms and assess performance by submitting the Docker container. Note that only five submission attempts are allowed during this period.

Final Test Phase

• Test Evaluation: Starting on July 15, 2024, the final test phase begins.
• Submission of Final Docker Container: Participants submit their final Docker container, which will be evaluated using a set of 40 unseen test images.
• Submission Attempts: During the final test phase, participants have two submission attempts to present their best-performing algorithm.

Evaluation Metrics

The submitted algorithms in the Aorta Segmentation Challenge will be evaluated using two key metrics:

1. Dice Coefficient (DSC):
• Measures the overlap between the predicted segmentation mask and the ground truth mask.
• Ranges from 0 (no overlap) to 1 (perfect overlap).
• The dice coefficient is computed for all 23 segmentation labels and then averaged. The reporting dice score represents the mean across all labels.
2. Normalized Surface Distance (NSD):
• Assesses the distance between the predicted surface and the ground truth surface.
• Ranges from 0 (maximum misalignment) to 1 (perfect alignment).
• The NSD is computed for all 23 segmentation labels and then averaged. The reporting NSD represents the mean across all labels for each subject.

Both metrics are crucial for evaluating the accuracy and robustness of segmentation algorithms. Participants should aim for high DSC values (close to 1) and higher NSD values (close to 1) to demonstrate the effectiveness of their models in accurately delineating the Aorta and related structures. Best of luck to all participants!