Related Experiment Video
Updated: Oct 4, 2025
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Radiologist-Level Performance by Using Deep Learning for Segmentation of Breast Cancers on MRI Scans
Lukas Hirsch1, Yu Huang1, Shaojun Luo1
1Department of Biomedical Engineering (L.H., Y.H., L.C.P.) and the Benjamin Levich Institute and Department of Physics (S.L., H.A.M.), the City College of the City University of New York, 160 Convent Ave, New York, NY 10031; Department of Radiology, Memorial Sloan Kettering Cancer Center, New York, NY 10065 (Y.H., C.R.S., R.L.G., I.D.N., A.G.V.B., N.O., E.S.K., D.L., D.A., S.E.W., M.H., D.F.M., K.P., K.J., A.E.E., P.E., E.A.M., E.J.S.); Department of Imaging, A.C. Camargo Cancer Center, São Paulo, Brazil (A.G.V.B.); Department of Radiology, University of California, San Francisco, San Francisco, Calif (N.O.); Department of Radiology, Samsung Medical Center, Sungkyunkwan University School of Medicine, Seoul, Korea (E.S.K.); and Department of Breast Imaging, Breast Cancer Center TecSalud, ITESM Monterrey, Monterrey, Mexico (D.A.).
A novel 3D U-Net deep learning model achieved radiologist-level accuracy in segmenting breast cancers on MRI scans. This supervised learning approach demonstrates the potential of artificial intelligence in improving breast cancer detection and diagnosis.
Area of Science:
- Medical Imaging and Radiology
- Artificial Intelligence in Healthcare
- Machine Learning for Medical Diagnosis
Background:
- Accurate segmentation of breast cancers in MRI is crucial for diagnosis and treatment planning.
- Current segmentation methods often rely on manual delineation by radiologists, which can be time-consuming and subject to inter-observer variability.
Purpose of the Study:
- To develop and evaluate a deep learning network for fully automated, radiologist-level segmentation of breast cancers using MRI.
- To compare the performance of the developed deep network against human radiologists in segmenting breast cancers.
Main Methods:
- A retrospective study utilized a large dataset of 38,229 MRI examinations (14,475 patients) including 2,555 segmented breast cancers and 60,108 benign breasts for training.
- A 3D U-Net architecture with dynamic contrast-enhanced MRI as input and intensity normalization was selected and trained.
- Performance was evaluated using the Dice score for 2D segmentation, comparing the network's results against independent segmentations by four radiologists on 250 breast cancers.
Main Results:
- The optimized 3D U-Net achieved a median Dice score of 0.77 on the test set.
- The deep learning network demonstrated performance equivalent to that of fellowship-trained radiologists, with a Dice score range of 0.69-0.84.
- The study highlights the efficacy of supervised learning and convolutional neural networks (CNNs) in medical image analysis.
Conclusions:
- A 3D U-Net, trained on a substantial dataset, can achieve radiologist-level performance for breast cancer segmentation in routine clinical MRI.
- This deep learning approach offers a promising tool for automating cancer segmentation, potentially enhancing diagnostic efficiency and accuracy.
- The findings support the integration of advanced machine learning algorithms into breast MRI interpretation workflows.
