T20: Generating caption from WSI¶

Objective:
The objective of this task is to generate a caption that encapsulates the relevant clinical information such as tissue type and diagnosis from whole-slide images (WSI). The WSI corresponds to a tissue sample obtained from the colon, rectum or cervix.

Patient Population:
The WSI-report pairs are obtained from the ExaMode project, which provides a large dataset ~10k comprising colon and cervix biopsies and polyps and their corresponding Dutch pathology reports.

Image-Report Data:

• A subsample of the initial dataset was selected to curate the challenge dataset for this task.
• Given that most of the biopsies are benign (more than half), a balanced selection of cases across each diagnostic category was made to ensure diversity within the dataset.
• The reports have been preprocessed to retain only the conclusion section.

Test Data:
The test data is the WSI without the pathology report. Participants must generate a caption as a string in Dutch that encapsulates the relevant clinical information such as tissue type and diagnosis from WSI.

Reference Standard:
The reference standard consists of the conclusion section from the original Dutch pathology report associated with each WSI, as it captures the pathologist’s final diagnostic interpretation.

Evaluation Metrics:
The generated caption is evaluated by comparing its similarity to the original pathology report using the following metrics:

• BLEU-4
• ROUGE-L
• CIDEr
• METEOR
• BERTScore F1

The final score is the average of all five metrics.