dataset

Hello, I see that your dataset is not registered, why can the ssim index be calculated?

I guess they may use structural_similarity from skimage.metrics However, SSIM and many other figures of merit in this challenge are not suitable for ultrasound imaging. Those figures of merit were designed for natural images, with a quite different statistics in comparison with clinical ultrasound imaging. That's why I'm seriusly considering not to participate in the challenge, as far as the way to measure how we enhance the images is just wrong. A pristine image, with soft texture and almost without noise ort other artefacts, is useless from the clinical point of view. We should work from accoustic signal to improve the reconstruction, not from already reconstructed images to avoid destroying relevant clinical information.

I have not downloaded the dataset, because then I shall be force to send a development which can be tricked to get high scores in useless "enhanced" images. As far as I can see (without downloading the dataset), the challenge is not well designed.

Thank you for your reminding. The training data is not strictly pixel aligned, but the registration preprocessing, including a B-spline-based nonrigid registration and a landmark-based nonrigid registration, has been adopted to obtain paired testing data for evaluating the reconstruction performance.

USenhance2023 Orgnizers

Thank you for clearing my doubts.