COSAS2024 comprises two tasks: cross-organ adenocarcinoma segmentation and cross-scanner adenocarcinoma segmentation. While the datasets differ between these tasks, the process, data volume, annotation characteristics, and evaluation metrics remain consistent.

🗂️Datasets¶

COSAS offers participants the first and currently largest domain-generalization dataset for digital pathology segmentation tasks. It introduces domain shift challenges arising from two distinct factors for the first time: different organs in Task 1 and varying scanners in Task 2.

💡Task 1: A total of 290 patch images of six different adenocarcinomas are used in this task. These images, with an average size of 1500x1500 pixels, are all extracted from WSIs digitised using the TEKSQRAY SQS-600P scanner.

The training set comprises images from 3 different organs (gastric adenocarcinoma, colorectal adenocarcinoma, and pancreatic ductal adenocarcinoma), with 60 images per type and 180 images in total; the preliminary test set consists of images from 4 different organs (including two from the training set), with 5 images per type and 20 images in total; the final test set contains images from all 6 organ types, with 15 images per type and 90 images in total.

💡Task 2: The data for Task 2 comprises patch images of Invasive breast carcinoma of no special type, acquired from six distinct WSI scanners, totalling 290 images. Each image has an approximate size of 1500x1500 pixels.

The training set includes images from 3 different scanners (TEKSQRAY SQS-600P, KFBIO KF-PRO-400 and  3DHISTECH PANNORAMIC 1000), with 60 images per type and 180 images in total; The preliminary test set comprises images from 4 different scanners, with 5 images per type and 20 images in total; The final test set includes images from all 6 scanners, with 15 images per type and 90 images in total.

Note: No additional data allowed. Pre-trained networks are permissible only if they have been trained on conventional, non-medical image datasets such as ImageNet or MS COCO.

📜Challenge Phases¶

The challenge consists of three phases:

• Training Phase (Duration: 2 months. No submission)

To participate in the COSAS challenge, teams must register at cosas.grand-challenge.org. During the training phase, participants can utilize the Public Training Set to develop and train their machine learning models.

• Preliminary Test Phase (Duration: 2 weeks. Submissions allowed: 5 total)

Each team can submit one trained algorithm in a Docker container for evaluation, with a maximum of 5 submissions per task. These algorithms undergo testing on the grand-challenge.org platform using a hidden preliminary test set. Results are displayed on a real-time, public leaderboard to maintain confidentiality and ensure the integrity of AI predictions.

• Final Test Phase (Duration: 3 days. Submissions allowed: 1 total)

Following the Preliminary Test Phase, teams can submit their best AI algorithm for final evaluation on a hidden Final Test Set. During the Final Test Phase, teams are permitted to submit their final AI algorithm only once. Evaluation on this set will determine the final rankings, identifying the top 5 best-performing models.
Throughout both the Preliminary Test and Final Test phases, algorithms will be executed on the grand-challenge.org platform. The ranking methodology assigns a weight of 0.2 to the Preliminary Test Set and 0.8 to the Final Test Set.