Gallagher-Syed A., Pontarini E., Bombardieri M, Lewis M. J., Slabaugh G., Barnes M., "Histopathological Assessment of Sjogren's disease with HistoMIL", IEEE Internal Symposium on Biomedical Imaging, Cartagena de Indias. 2023. Conference abstract.

A modified CLAM/MIL pipeline for large histopathology WSIs, composed of:

1. A VGG16 embedding backbone, previously trained on the WSI patches with labels propagated from the slide level, reducing each patch to a 1024 feature vector.
2. All embeddings from a slide are aggregated into a larger feature vector and passed to the attention model
3. The attention modell (CLAM/MIL) aggregates the patch-level information to the slide level, as well as calculating an attention score for each patch
4. A slide level prediction is made
5. A heatmap is created for the slide

References:
https://github.com/mahmoodlab/CLAM#readme
https://github.com/AMLab-Amsterdam/AttentionDeepMIL