Patch level features and MIL pooling

Hi Unicorn Team,

I recently watched the kick-off video and, it’s stated that features must be constrained to case level and patch level features aren’t allowed for WSI classification or regression (kick-off video). However, on the forum, participants have shared patch level feature files. For example, this one: feature file (see discussion here: discussion)

Could you please clarify whether the original restriction on patch level features still applies? If patch level features are now allowed, would using an adaptor with MIL pooling be within the challenge guidelines?

Thank you!

Best,

Michel

Hi Michel,

happy to clarify!

For classification and regression tasks, features must be provided at the case level. This is because the labels are also defined at the case level, and allowing patch-level features would enable the use of more complex adaptors (e.g., MIL models), which we want to avoid. One of our goals is to ensure that the adaptors remain relatively simple so that model performance more directly reflects the quality of the foundation model's features, rather than the complexity of the downstream model.

Additionally, in the context of the challenge, where only 48 shots are provided to train the adaptor for vision tasks, training a complex model like a MIL aggregator on patch-level features would be very challenging.

However, for dense prediction tasks (segmentation and detection), patch-level features are allowed and expected, since the labels are provided at a finer spatial resolution (typically pixel-level), and case-level features alone wouldn't capture the necessary detail to perform the task. That's why patch-level features are allowed for Task 5.

clément.

Hi Clement,

Perfectly clear, thank you!

Best, Michel