T9: Segmenting ROIs in Breast Cancer H\&E-Stained WSIs¶
Objective: Segment region-of-interest (ROI) cropouts from H&E-stained whole-slide images (WSIs) of breast tissue into three tissue classes: tumor (1), stroma (2), and other (3), with background labeled as 0.
Patient Population: Patients with breast cancer from Radboudumc and the Jules Bordet Institute, with one WSI per patient.
Imaging Data:
Multi-resolution H&E-stained ROI cropouts extracted from WSIs. Images are provided as .tif files, approximately 1000×1000 pixels in size, with a resolution of 0.5 microns per pixel.
Reference Standard: Annotations were made by five board-certified breast pathologists from the Tumor Infiltrating Lymphocyte (TIL) Working Group, based on predefined ROIs.
Evaluation Metrics: Average DICE coefficient of the tumor and stroma classes. DICE is computed per class and then averaged.
Relation to Existing Challenges:¶
- Task 9 is derived from the TIGER challenge.
- Unlike in the TIGER challenge, which focused on training task-specific models locally using a large labeled dataset, UNICORN shifts training to the platform, using a small set of few-shot examples per task.
- Validation and test datasets are subsets of the original TIGER dataset.
- The original seven annotated tissue classes from TIGER were merged into three:
- Invasive tumor and in-situ tumor → tumor
- Tumor-associated stroma and inflamed stroma → stroma
- Healthy glands, necrosis not in-situ, and rest → rest