T5: Cell detection of signet ring cells in H&E-stained WSI of gastric cancer¶

Objective:
Develop a model to detect individual signet ring cells in H&E-stained regions of interest (ROIs) extracted from whole slide images (WSIs) of gastric cancer. The goal is to accurately localize these rare tumor cells by predicting their (x, y) coordinates in the image.

Patient Population:
35 patients diagnosed with gastric cancer at Radboudumc. All available tumor regions, including pre-invasive lesions, were considered.

Imaging Data:

• 259 H&E-stained WSIs, scanned using the same scanner
• ROIs were extracted from these WSIs and saved as individual .tif files
• Resolution: 0.5 microns per pixel
• Annotations provided in .json format, with each file containing a list of (x, y) coordinates marking the center of each annotated signet ring cell

Test Data:
Unlabeled ROI images (.tif format) at the same resolution, without annotations. The model's output should be a JSON file containing the image coordinates (x, y) of detected signet ring cells.

Reference Standard:

• ROIs were selected by a resident pathologist focusing on diffuse gastric cancer lesions, including pre-invasive areas
• Signet ring cells within these ROIs were manually annotated by trained student annotators, with coordinates representing the center of each cell

Evaluation Metrics:
Model performance will be evaluated using the F1-score based on the spatial match between predicted and reference coordinates.