A Deep Learning-based Model for Detecting Abnormalities on Brain MR Images for Triaging: Preliminary Results from a
Romane Gauriau1, Bernardo C Bizzo1, Felipe C Kitamura1
1MGH & BWH Center for Clinical Data Science, Ste 1303, Floor 13, 100 Cambridge St, Boston, MA 02114 (R.G., B.C.B., F.B.C.M., K.P.A.); Department of Artificial Intelligence, Diagnósticos da América, São Paulo, Brazil (B.C.B., F.C.K., O.L.J., S.F.F., M.R.T.G., L.M.V., R.C.D., E.L.G.); Head of AI, Diagnósticos da América SA, São Paulo, Brazil (F.C.K.); Department of Radiology, Federal University of Rio de Janeiro, Rio de Janeiro, Brazil (B.C.B., T.A.S., E.L.G.); Department of Radiology, Massachusetts General Hospital, Boston, Mass (B.C.B.); and Department of Radiology, Brigham and Women's Hospital and Harvard Medical School, Harvard University, Boston, Mass (K.P.A.).
A deep learning model effectively detects brain abnormalities on MR images. This convolutional neural network shows good performance in differentiating normal from abnormal findings across institutions.
Area of Science:
- Medical Imaging
- Artificial Intelligence in Medicine
- Neurology
Background:
- Brain MR imaging is crucial for diagnosing neurological conditions.
- Accurate interpretation of MR images is vital for patient care.
- Developing automated tools can aid radiologists in detecting abnormalities.
Purpose of the Study:
- To develop a deep learning model for detecting brain abnormalities on MR images.
- To classify brain MRI findings as 'likely normal' or 'likely abnormal' using T2-weighted fluid-attenuated inversion recovery images.
- To evaluate the model's performance on heterogeneous datasets from different institutions.
Main Methods:
- A retrospective study utilizing a convolutional neural network (CNN) model.
- Training on large, multi-continental datasets (Datasets A and B) covering various pathologies.
- Testing models on independent datasets (subsets of A, B, and Dataset C) and comparing with radiology report annotations.
Main Results:
- Model A, trained on one institution's data and tested on another's (Dataset C), achieved an F1 score of 0.72.
- The same model (Model A) obtained an area under the receiver operating characteristic curve of 0.78 when compared to radiology reports.
- Performance was evaluated against both image-based annotations and radiology report labels.
Conclusions:
- The developed deep learning model demonstrates good performance in distinguishing between normal and abnormal brain MRIs.
- The model's ability to generalize across different institutions highlights its potential clinical utility.
- This AI approach shows promise for supporting the interpretation of brain MR imaging studies.
