Using CHIEF

[suminwei]

Hello,
I have a more naive question regarding using CHIEF
I was wondering is CHIEF mainly intended to be used as an encoder, and need to be used with a classification algorithm which takes the extracted features as input? If this is the case, what classification algorithm did you use when you performed the cancer detection task and compared to other weakly supervised methods (ie CLAM, ABMIL and DSMIL)?
Or, can CHIEF be used as a standalone to classify pathology images into different classes?
Thank you!
Lisa

[Xiyue-Wang]

Thank you, Lisa. Yes, your understanding is mostly correct. CHIEF can indeed be used for extracting feature representations from H&E-stained histological images. In our paper, we first employed the CHIEF architecture for a two-class classification task (i.e., cancer or non-cancer), and we fine-tuned it for other tasks. If you're aiming to perform a classification task with more classes, you can either freeze the model and extract WSI-level features for your classifiers or fine-tune the CHIEF model to suit your specific needs. In either case, be sure to replace the classifier with one designed for multi-class classification. Please feel free to let us know if you have more questions.