Unable to get the model

[drashokpuri]

I couldn't find this file. ./RETFound_cfp.pth', and would like to know how to download the checkpoint.

[drashokpuri]

Sorry for the trouble. I found it. But what I get on reconstruction is

[drashokpuri]

What I get with the facebook/mae is:

[rmaphoh]

Thanks for your feedback. Can I check if 1) the checkpoint you used is the OCT version and 2) the two original inputs are identical (they look quite different in contrast according to the screenshots you showed)?

The blueish patches occur when the --norm_pix_loss is set as True (in default). Each patch is normalised individually.

[drashokpuri]

Hi, Thanks for your  reply.I have used the checkpoint of the OCT version and not the retinal fundus version.I too noted that the contrast in the 2 images is different despite the same code used for the input image in both cases. It is a very  nice concept of yours to create a retinal foundation model in lines of a large language foundation model. I have also removed the decoder from your model and replaced with a classifier to perform  classification  task. I have used a subset of the Kermany dataset of OCT images for training. I used the same test set of 968 images for testing, but could not get a high AUC-ROC like what you have got.Maybe I need some inputs from you for this. Best regards,Ashok

On Thursday, October 5, 2023 at 08:30:28 PM GMT+5:30, Yukun Zhou ***@***.***> wrote:  

Thanks for your feedback. Can I check if 1) the checkpoint you used is the OCT version and 2) the two original inputs are identical (they look quite different in contrast according to the screenshots you showed)?

The blueish patches occur when the --norm_pix_loss is set as True (in default). Each patch is normalised individually.

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

[rmaphoh]

We used the Kermany dataset for self-supervised learning, rather than downstream task evaluation.

We do not suggest evaluating the disease detection on the Keymany dataset as the images have been seen by models (not for disease labels though).

[drashokpuri]

Thanks for your response.I used the test set of Kermany et al for evaluation with the presumption that these images were not shown to the model during training.Can you please suggest a test dataset for evaluation of the autoencoder turned classifier for retinal OCT images that you have used.Regards

On Monday, October 16, 2023 at 04:32:17 AM GMT+5:30, Yukun Zhou ***@***.***> wrote:  

We used the Kermany dataset for self-supervised learning, rather than downstream task evaluation.

We do not suggest evaluating the disease detection on the Keymany dataset as the images have been seen by models (not for disease labels though).

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

[rmaphoh]

In this paper, we used the OCTID (https://borealisdata.ca/dataverse/OCTID) for the evaluation.

[s20488]

I apologise but I'm having the same problem, I can't find the ‘./RETFound_cfp.pth’ file, could you please advise?