Share Email Print
cover

Proceedings Paper • new

Application of deep learning techniques to characterization of 3D radiological datasets: a pilot study for detection of intravenous contrast in breast MRI (Conference Presentation)
Author(s): Krishna Nand Keshavamurthy; Pierre Elnajjar; Amin El-Rowmeim; Hao-Hsin Shih; Ian Pan; Kinh Gian Do; Krishna Juluru
cover GOOD NEWS! Your organization subscribes to the SPIE Digital Library. You may be able to download this paper for free. Check Access

Paper Abstract

Radiological images are stored using the DICOM standard with metadata that includes patient identifiers and acquisition parameters. Some parameters are inherent to the settings on a scanner, such as tube current in CT or echo time (TE) in MRI. These are reliably recorded into the metadata. Other parameters, however, are not inherent to the scanner settings, and therefore require user input which is prone to human error. For clinical and research purposes, there is a general need for automated approaches that will appropriately classify images, even with parameters that are not inherent to the scanner settings. In this work, we present a deep learning based approach for automatically detecting one such parameter: the absence or presence of intravenous contrast in MRI scans. Our classifier is a convolutional neural network [1] based on the ResNet [2] architecture, which is popular in computer vision for various tasks such as object detection, recognition etc. Our data consisted of 1000 randomized breast MRI scans (500 scans with and 500 scans without intravenous contrast), acquired at the Memorial Sloan Kettering Cancer Center. The labels for the scans were obtained from the series descriptions that contained pertinent words such as “pre” and “post”. We reserved 80% of the data for training the CNN and 20% for testing. Preliminary results are very promising (area under the ROC curve = 0.98; optimal cut-off point (maximal Youden’s index [3]) sensitivity = 1.0, specificity = 0.96), indicating the potential usefulness of this technique in clinical and research scenarios.

Paper Details

Date Published: 15 March 2019
PDF
Proc. SPIE 10954, Medical Imaging 2019: Imaging Informatics for Healthcare, Research, and Applications, 109540X (15 March 2019); doi: 10.1117/12.2513809
Show Author Affiliations
Krishna Nand Keshavamurthy, Memorial Sloan-Kettering Cancer Ctr. (United States)
Pierre Elnajjar, Memorial Sloan-Kettering Cancer Ctr. (United States)
Amin El-Rowmeim, Memorial Sloan-Kettering Cancer Ctr. (United States)
Hao-Hsin Shih, Memorial Sloan-Kettering Cancer Ctr. (United States)
Ian Pan, Brown Univ. (United States)
Kinh Gian Do, Memorial Sloan-Kettering Cancer Ctr. (United States)
Krishna Juluru, Memorial Sloan-Kettering Cancer Ctr. (United States)


Published in SPIE Proceedings Vol. 10954:
Medical Imaging 2019: Imaging Informatics for Healthcare, Research, and Applications
Po-Hao Chen; Peter R. Bak, Editor(s)

© SPIE. Terms of Use
Back to Top