Welcome to K2S: from undersampled K-space to automatic Segmentation

Magnetic resonance imaging (MRI) is the modality of choice for evaluating knee joint degeneration, but it can be susceptible to long acquisition times, tedious post processing, and lack of standardization. One of the most compelling applications of deep learning, therefore, is accelerated analysis of knee MRI. In addition to faster MRI acquisition, deep learning has enhanced image post-processing applications such as tissue segmentation.

While fast, undersampled MRI acquisition may not have qualitative, visual acuity that comes from fully-sampled data, the underlying embedding space may be adequate for some applications. The implications for down-stream tasks such as tissue segmentation using convolutional neural networks are not well-characterized. Efficient segmentation of key anatomical structures from undersampled data is an open question that has clinical relevance, e.g., patient triage. The goal of this challenge, therefore, is to train segmentation models from 8x undersampled knee MRI.

This challenge consists of a dataset of high-resolution 3D knee MRI including raw k-space data and post-processing annotations with masks for tissue segmentation. The 8x undersampled , multi-channel k-space data of 300 fat-suppressed 3D FSE Cube sequences along with segmentation masks of anatomical structures (cartilage, bone) will be shared. Fully-sampled images will also be shared for training as auxiliary information, but will not be available for inference on the test set.

Submissions will be evaluated in an end-to-end fashion from undersampled k-space data to segmentation on a test set of 50 fat-suppressed FSE-Cube sequences (undersampling pattern same as training data).

Top submissions will be invited to contribute to a full-length challenge summary research article.