The Stanford ML Group led by Andrew Ng has released its MRNet Dataset, which contains more than 1000 annotated knee MRI Scans; and announced an associated public model competition.
The MRNet Dataset contains data from 1,370 knee MRI (magnetic resonance imaging) exams performed at the Stanford University Medical Center between January 1, 2001 and December 31, 2012. Of these, 1,104 (80.6%) are abnormal cases, with 319 (23.3%) diagnosed as ACL (anterior cruciate ligament) tears and 508 (37.1%) as Meniscal tears.
For model training, the dataset is further divided into a training set with 1,130 exams, a validation set with 120 exams, and a hidden test set with 120 exams. The MRNet Dataset was originally used to develop MRNet, a deep learning model that can “rapidly generate accurate clinical pathology classifications of knee MRI exams.” Concurrent with open-sourcing the dataset, Stanford ML Group announced an open competition for models which can automatically interpret the knee MRIs.
Researchers and interested parties can register at this link to download the dataset and participate in the competition. Submitted algorithms will run on the hidden test set, and the average AUC for abnormality detection of ACL or Meniscal tears will be used as the measurement metric.
Trained on the MRNet dataset, MRNet is a CNN-based model that maps a 3-dimensional MRI series to a probability, in order to predict abnormalities in knee MRI exams.
The trained prediction algorithm’s performance was tested in real-world scenarios to help radiologists and surgeons with diagnosis. The model’s use significantly reduced false positive rates wherein abnormalities were mistakenly diagnosed in healthy patients.
More detailed information on MRNet is available at PLOS Medicine.
Author: Mos Zhang | Editor: Michael Sarazen