Patch feature extraction with Titan

[JAAY-AI]

In the builder.py file, the code for loading the Titan model is as follows:

        model, _ = titan.return_conch()
        assert target_img_size == 448, 'TITAN is used with 448x448 CONCH v1.5 features'

However, the Titan official documentation states ‘TITAN-preview is a vision-lanuage model trained on CONCH v1.5 patch features with patch size of 512x512 pixels at 20x magnification.’.

Why is there this discrepancy? Furthermore, will the feature extraction results be significantly worse if I perform it on 224x224px patches at a 5x magnification

[guillaumejaume]

Please refer to our new repo trident for titan/conchv1.5 feature extraction: https://github.com/mahmoodlab/trident.

However, the Titan official documentation states ‘TITAN-preview is a vision-lanuage model trained on CONCH v1.5 patch features with patch size of 512x512 pixels at 20x magnification.’.
==> it's 512x512 at 20x resized to 448x448.

Why is there this discrepancy? Furthermore, will the feature extraction results be significantly worse if I perform it on 224x224px patches at a 5x magnification
==> likely yes.

[JAAY-AI]

Thank you for your answer！

…

---- Replied Message ---- | From | Guillaume ***@***.***> | | Date | 04/03/2025 22:09 | | To | mahmoodlab/CLAM ***@***.***> | | Cc | LiuJin ***@***.***>, Author ***@***.***> | | Subject | Re: [mahmoodlab/CLAM] Patch feature extraction with Titan (Issue #302) | Please refer to our new repo trident for titan/conchv1.5 feature extraction: https://github.com/mahmoodlab/trident. However, the Titan official documentation states ‘TITAN-preview is a vision-lanuage model trained on CONCH v1.5 patch features with patch size of 512x512 pixels at 20x magnification.’. ==> it's 512x512 at 20x resized to 448x448. Why is there this discrepancy? Furthermore, will the feature extraction results be significantly worse if I perform it on 224x224px patches at a 5x magnification ==> likely yes. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***> guillaumejaume left a comment (mahmoodlab/CLAM#302) Please refer to our new repo trident for titan/conchv1.5 feature extraction: https://github.com/mahmoodlab/trident. However, the Titan official documentation states ‘TITAN-preview is a vision-lanuage model trained on CONCH v1.5 patch features with patch size of 512x512 pixels at 20x magnification.’. ==> it's 512x512 at 20x resized to 448x448. Why is there this discrepancy? Furthermore, will the feature extraction results be significantly worse if I perform it on 224x224px patches at a 5x magnification ==> likely yes. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>