fastMRI

Annotated pathologies in the fastMRI knee dataset

Dear users of the fastMRI dataset,

Over the last couple of months, I received several requests from researchers who followed the fastMRI challenge, to point them to the files in the dataset that correspond to the missed pathology that we show in the challenge paper, in particular the one in Figure 3: https://onlinelibrary.wiley.com/doi/full/10.1002/mrm.28338. They wanted to use these examples to test their own algorithms with respect to preservation of the pathology. I have so far declined these requests because we have decided very early on that we will not release the challenge dataset to the public after the conclusion of the challenge, because we don’t want people to overfit on it, in particular since everybody now knows the correct ground truth.

However, since one of our messages from the challenge has been that looking at pathology is important instead of just focusing on metrics like SSIM, we wanted to support these type of experiments. Together with a member of our MSK team at NYU radiology, Dr. Ben Wang (https://nyulangone.org/doctors/1932441029/benjamin-wang), we identified some cases in the public training set with comparable pathology to what we show in the paper. Two of those cases are from 3T systems, one is from a 1.5T system, so they provide a different SNR baseline to start with. The attached Figures show the corresponding cases slices with annotations that point towards the meniscal tears.

Case 1: file 1000425: 1.5T (Siemens Aera): Slice 23: Tear of medial meniscus, potential tear of lateral meniscus.

Case 2: file 1000823: 3T (Siemens Skyra): Slice 24: Tear of medial meniscus

Case 3: file 1002455: 3T (Siemens Skyra): Slice 27: Tear of medial and lateral meniscus

Just to let you know, this is essentially a preview. We are currently working on creating a full set of segmentations of pathology for a subset of the knee dataset, which we are planning to provide to the research community. I hope that this further enhances the usefulness of the dataset to the research community. In the meantime, I hope that this first glimpse at the meniscal tears is already useful for your work.

Best wishes, Florian

2 Likes