Skip to content

Insufficient H&E resolution for meaningful H&E-ST contrastive learning #56

@NBitBuilder

Description

@NBitBuilder

ST-Bank data: most H&E images are around $16 \times 16$

Image Image Image

After reviewing the ST-Bank dataset, I noted that most H&E images are cropped to approximately $16 \times 16$ pixels (matching the 55µm ST spot size). At this resolution, individual cells are reduced to just a few pixels, resulting in a total loss of the morphological detail necessary for contrastive learning. How can a model effectively learn meaningful biological representations when the input data lacks cellular features? This seems to be a fundamental bottleneck that undermines the validity of the H&E-ST training approach.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions