arXiv: Computer Vision and Pattern Recognition | , Vol arXiv:2109.03812
Improving speed and image quality of Magnetic Resonance Imaging (MRI) via novel reconstruction approaches remains one of the highest impact applications for deep learning in medical imaging. The fastMRI (opens in new tab) dataset and associated reconstruction challenges have been key drivers in the development of machine learning based image reconstruction techniques.
The fastMRI is one of the largest collections of MRI raw data (fully sampled k-space data) and has enabled users without direct access to imaging systems to develop deep learning-based reconstruction methods. While the impact of the fastMRI dataset on the field of medical imaging is unquestioned, the dataset currently lacks clinical expert pathology annotations, critical to addressing clinically relevant reconstruction frameworks and exploring important questions regarding rendering of specific pathology using such novel approaches.
This work introduces fastMRI+ (opens in new tab), which consists of 16154 subspecialist expert bounding box annotations and 13 study-level labels for 22 different pathology categories on the fastMRI knee dataset, and 7570 subspecialist expert bounding box annotations and 643 study-level labels for 30 different pathology categories for the fastMRI brain dataset.
The fastMRI+ dataset is open access and aims to support further research and advancement of medical imaging in MRI reconstruction and beyond. To access the dataset labels and instructions for using the labels, see https://github.com/microsoft/fastmri-plus (opens in new tab).
The purpose of this project is to release a set of clinically relevant labels for an existing, publicly available dataset (https://fastmri.org). With permission from the fastMRI team at NYU, we have engaged third party labelers (radiologist) to place bounding boxes on lesions in the fastMRI dataset and we would like to release those labels to the community. We intend to release the labels (CSV files) on GitHub.
