Testing a new image

[maybefreedom]

Hello! I've read your code + article and find this project extremely interesing. I think this might be obvious but I have trained my datasets and everything, now I need to test a new fundus image to see if the MA are detected. Is there a testing method I've overlooked?

Thank you and sorry again if the answer is obvious.

[ignaciorlando]

Hi @maybefreedom!

Thanks for your interest in our paper and code.

Yes, you can use this function: [red_lesion_segmentation, score_map] = full_red_lesion_segmentation(I, mask, L0, step, L, K, px, cnn_for_feature_extraction, vessel_segmentation, detector)

This parameters are:

• I: the input image.
• mask: the FOV mask (you can use mask = get_fov_mask(I, threshold) to get it, try with different thresholds or use the values we reported in the paper).
• L0: the lowest value of L (you can set it in 3 as in our paper).
• step: the step to generate the L scales (you can set it in 3 as in our paper).
• L: the highest value of L (you can set it in 60 as in our paper).
• K: the maximum number of candidates (you can set it in 120 as in our paper).
• px: structures with less than px pixels will be removed (you can set it in 5 as in our paper).
• cnn_for_feature_extraction: the CNN structure to retrieve the features (you can use the CNN that you trained or load one of our pretrained networks).
• vessel_segmentation: the vessel segmentation. You can compute it using our segmentation method, or use a logical matrix with the same size of the image but full of zeros to ignore the segmentation feature.
• detector: the RF classifier, trained using our hybrid approach. You can use the model you learned or the pre-trained classifier we provide.

Let me know if you were able to run it!

Cheers,

Nacho