TY - JOUR
T1 - Deep Learning-Based Automatic Detection of Brain Metastases in Heterogenous Multi-Institutional Magnetic Resonance Imaging Sets
T2 - An Exploratory Analysis of NRG CC001
AU - Liang, Ying
AU - Lee, Karen
AU - Bovi, Joseph A.
AU - Palmer, Joshua D.
AU - Brown, Paul D.
AU - Gondi, Vinai
AU - Tomé, Wolfgang A.
AU - Benzinger, Tammie L.S.
AU - Mehta, Minesh P.
AU - Li, X. Allen
N1 - Funding Information:
This work was partially supported by the National Cancer Institute of the National Institutes of Health under Award Numbers UG1CA189867 and U24CA180803. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.
Funding Information:
Inputs and discussions from Markus Sprenger, George Noid, PhD, Nguyen Phuong Dang, PhD, and Manav Bhalla, MD are appreciated. This work was partially supported by the National Cancer Institute of the National Institutes of Health under Award Numbers UG1CA189867 and U24CA180803. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. Disclosures: J.D.P. reports pediatric clinical trial funding from the National Institutes of Health (NIH) R702, personal fees for Huron Consulting Group and MORE Health, payment for advisory board for Varian Medical Systems and Novocure, and support for attending meetings or travel for NovoCure. P.D.B. reports honorarium to self as a contributor to UpToDate (unrelated to current topic). W.A.T. reports research support from Accuray and Varian to institution, research support payments from the Wisconsin Alumni Research and consulting fees from Accuray, payment for Data Safety Monitoring Board, and Advisory Board for Archeus. T.L.S.B. reports grants or contracts for Alzeimer's Association from NIH, Advisory Board and consulting fees from Biogen, unpaid consultant for Eisai and Siemens, participation on a Data Safety Monitoring Board or Advisory Board for the NIH, receipt of materials from Avid Radiopharmaceuticals (precursor and tech transfer), Cerveau (precursor and tech transfer), and Life Molecular Imaging (precursor and tech transfer). M.P.M. reports consulting fees for Mevion, Zap, Sapience, Blue Earth Diagnostics, IBA, and Xoft, NRG Oncology leadership role, stock or stock options for Chimerix, and Board of Directors with Oncoceutics. X.A.L. reports research grants from Elekta, Manteia, Siemens, and Accuray to his institution, Honoraria for educational lectures from Elekta and Accuray. The remaining authors have no conflicts of interest to declare.
Funding Information:
Disclosures: J.D.P. reports pediatric clinical trial funding from the National Institutes of Health (NIH) R702, personal fees for Huron Consulting Group and MORE Health, payment for advisory board for Varian Medical Systems and Novocure, and support for attending meetings or travel for NovoCure. P.D.B. reports honorarium to self as a contributor to UpToDate (unrelated to current topic). W.A.T. reports research support from Accuray and Varian to institution, research support payments from the Wisconsin Alumni Research and consulting fees from Accuray, payment for Data Safety Monitoring Board, and Advisory Board for Archeus. T.L.S.B. reports grants or contracts for Alzeimer's Association from NIH, Advisory Board and consulting fees from Biogen, unpaid consultant for Eisai and Siemens, participation on a Data Safety Monitoring Board or Advisory Board for the NIH, receipt of materials from Avid Radiopharmaceuticals (precursor and tech transfer), Cerveau (precursor and tech transfer), and Life Molecular Imaging (precursor and tech transfer). M.P.M. reports consulting fees for Mevion, Zap, Sapience, Blue Earth Diagnostics, IBA, and Xoft, NRG Oncology leadership role, stock or stock options for Chimerix, and Board of Directors with Oncoceutics. X.A.L. reports research grants from Elekta, Manteia, Siemens, and Accuray to his institution, Honoraria for educational lectures from Elekta and Accuray. The remaining authors have no conflicts of interest to declare.
Publisher Copyright:
© 2022 Elsevier Inc.
PY - 2022/11/1
Y1 - 2022/11/1
N2 - Purpose: Deep learning-based algorithms have been shown to be able to automatically detect and segment brain metastases (BMs) in magnetic resonance imaging, mostly based on single-institutional data sets. This work aimed to investigate the use of deep convolutional neural networks (DCNN) for BM detection and segmentation on a highly heterogeneous multi-institutional data set. Methods and Materials: A total of 407 patients from 98 institutions were randomly split into 326 patients from 78 institutions for training/validation and 81 patients from 20 institutions for unbiased testing. The data set contained T1-weighted gadolinium and T2-weighted fluid-attenuated inversion recovery magnetic resonance imaging acquired on diverse scanners using different pulse sequences and various acquisition parameters. Several variants of 3-dimensional U-Net based DCNN models were trained and tuned using 5-fold cross validation on the training set. Performances of different models were compared based on Dice similarity coefficient for segmentation and sensitivity and false positive rate (FPR) for detection. The best performing model was evaluated on the test set. Results: A DCNN with an input size of 64 × 64 × 64 and an equal number of 128 kernels for all convolutional layers using instance normalization was identified as the best performing model (Dice similarity coefficient 0.73, sensitivity 0.86, and FPR 1.9) in the 5-fold cross validation experiments. The best performing model demonstrated consistent behavior on the test set (Dice similarity coefficient 0.73, sensitivity 0.91, and FPR 1.7) and successfully detected 7 BMs (out of 327) that were missed during manual delineation. For large BMs with diameters greater than 12 mm, the sensitivity and FPR improved to 0.98 and 0.3, respectively. Conclusions: The DCNN model developed can automatically detect and segment brain metastases with reasonable accuracy, high sensitivity, and low FPR on a multi-institutional data set with nonprespecified and highly variable magnetic resonance imaging sequences. For large BMs, the model achieved clinically relevant results. The model is robust and may be potentially used in real-world situations.
AB - Purpose: Deep learning-based algorithms have been shown to be able to automatically detect and segment brain metastases (BMs) in magnetic resonance imaging, mostly based on single-institutional data sets. This work aimed to investigate the use of deep convolutional neural networks (DCNN) for BM detection and segmentation on a highly heterogeneous multi-institutional data set. Methods and Materials: A total of 407 patients from 98 institutions were randomly split into 326 patients from 78 institutions for training/validation and 81 patients from 20 institutions for unbiased testing. The data set contained T1-weighted gadolinium and T2-weighted fluid-attenuated inversion recovery magnetic resonance imaging acquired on diverse scanners using different pulse sequences and various acquisition parameters. Several variants of 3-dimensional U-Net based DCNN models were trained and tuned using 5-fold cross validation on the training set. Performances of different models were compared based on Dice similarity coefficient for segmentation and sensitivity and false positive rate (FPR) for detection. The best performing model was evaluated on the test set. Results: A DCNN with an input size of 64 × 64 × 64 and an equal number of 128 kernels for all convolutional layers using instance normalization was identified as the best performing model (Dice similarity coefficient 0.73, sensitivity 0.86, and FPR 1.9) in the 5-fold cross validation experiments. The best performing model demonstrated consistent behavior on the test set (Dice similarity coefficient 0.73, sensitivity 0.91, and FPR 1.7) and successfully detected 7 BMs (out of 327) that were missed during manual delineation. For large BMs with diameters greater than 12 mm, the sensitivity and FPR improved to 0.98 and 0.3, respectively. Conclusions: The DCNN model developed can automatically detect and segment brain metastases with reasonable accuracy, high sensitivity, and low FPR on a multi-institutional data set with nonprespecified and highly variable magnetic resonance imaging sequences. For large BMs, the model achieved clinically relevant results. The model is robust and may be potentially used in real-world situations.
UR - http://www.scopus.com/inward/record.url?scp=85136651527&partnerID=8YFLogxK
U2 - 10.1016/j.ijrobp.2022.06.081
DO - 10.1016/j.ijrobp.2022.06.081
M3 - Article
C2 - 35787927
AN - SCOPUS:85136651527
SN - 0360-3016
JO - International Journal of Radiation Oncology Biology Physics
JF - International Journal of Radiation Oncology Biology Physics
ER -