What does deep learning see? Insights from a classifier trained to predict contrast enhancement phase from CT images

Kenneth A. Philbrick, Kotaro Yoshida, Dai Inoue, Zeynettin Akkus, Timothy Kline, Alexander D. Weston, Panagiotis Korfiatis, Naoki M Takahashi, Bradley J Erickson

Research output: Contribution to journalArticle

7 Citations (Scopus)

Abstract

OBJECTIVE. Deep learning has shown great promise for improving medical image classification tasks. However, knowing what aspects of an image the deep learning system uses or, in a manner of speaking, sees to make its prediction is difficult. MATERIALS AND METHODS. Within a radiologic imaging context, we investigated the utility of methods designed to identify features within images on which deep learning activates. In this study, we developed a classifier to identify contrast enhancement phase from whole-slice CT data. We then used this classifier as an easily interpretable system to explore the utility of class activation map (CAMs), gradient-weighted class activation maps (Grad-CAMs), saliency maps, guided backpropagation maps, and the saliency activation map, a novel map reported here, to identify image features the model used when performing prediction. RESULTS. All techniques identified voxels within imaging that the classifier used. SAMs had greater specificity than did guided backpropagation maps, CAMs, and Grad-CAMs at identifying voxels within imaging that the model used to perform prediction. At shallow network layers, SAMs had greater specificity than Grad-CAMs at identifying input voxels that the layers within the model used to perform prediction. CONCLUSION. As a whole, voxel-level visualizations and visualizations of the imaging features that activate shallow network layers are powerful techniques to identify features that deep learning models use when performing prediction.

Original languageEnglish (US)
Pages (from-to)1184-1193
Number of pages10
JournalAmerican Journal of Roentgenology
Volume211
Issue number6
DOIs
StatePublished - Dec 1 2018

Fingerprint

Learning

Keywords

  • Class activation map (CAM)
  • Computer-aided diagnosis
  • Contrast enhancement phase
  • Convolutional neural network (CNN)
  • CT
  • Deep learning
  • Gradient-weighted class activation map (Grad-CAM)
  • Guided backpropagation
  • Machine learning
  • Saliency activation map
  • Saliency map

ASJC Scopus subject areas

  • Radiology Nuclear Medicine and imaging

Cite this

What does deep learning see? Insights from a classifier trained to predict contrast enhancement phase from CT images. / Philbrick, Kenneth A.; Yoshida, Kotaro; Inoue, Dai; Akkus, Zeynettin; Kline, Timothy; Weston, Alexander D.; Korfiatis, Panagiotis; Takahashi, Naoki M; Erickson, Bradley J.

In: American Journal of Roentgenology, Vol. 211, No. 6, 01.12.2018, p. 1184-1193.

Research output: Contribution to journalArticle

Philbrick, Kenneth A. ; Yoshida, Kotaro ; Inoue, Dai ; Akkus, Zeynettin ; Kline, Timothy ; Weston, Alexander D. ; Korfiatis, Panagiotis ; Takahashi, Naoki M ; Erickson, Bradley J. / What does deep learning see? Insights from a classifier trained to predict contrast enhancement phase from CT images. In: American Journal of Roentgenology. 2018 ; Vol. 211, No. 6. pp. 1184-1193.
@article{455c53b35a604ccb88075b80c28ab164,
title = "What does deep learning see? Insights from a classifier trained to predict contrast enhancement phase from CT images",
abstract = "OBJECTIVE. Deep learning has shown great promise for improving medical image classification tasks. However, knowing what aspects of an image the deep learning system uses or, in a manner of speaking, sees to make its prediction is difficult. MATERIALS AND METHODS. Within a radiologic imaging context, we investigated the utility of methods designed to identify features within images on which deep learning activates. In this study, we developed a classifier to identify contrast enhancement phase from whole-slice CT data. We then used this classifier as an easily interpretable system to explore the utility of class activation map (CAMs), gradient-weighted class activation maps (Grad-CAMs), saliency maps, guided backpropagation maps, and the saliency activation map, a novel map reported here, to identify image features the model used when performing prediction. RESULTS. All techniques identified voxels within imaging that the classifier used. SAMs had greater specificity than did guided backpropagation maps, CAMs, and Grad-CAMs at identifying voxels within imaging that the model used to perform prediction. At shallow network layers, SAMs had greater specificity than Grad-CAMs at identifying input voxels that the layers within the model used to perform prediction. CONCLUSION. As a whole, voxel-level visualizations and visualizations of the imaging features that activate shallow network layers are powerful techniques to identify features that deep learning models use when performing prediction.",
keywords = "Class activation map (CAM), Computer-aided diagnosis, Contrast enhancement phase, Convolutional neural network (CNN), CT, Deep learning, Gradient-weighted class activation map (Grad-CAM), Guided backpropagation, Machine learning, Saliency activation map, Saliency map",
author = "Philbrick, {Kenneth A.} and Kotaro Yoshida and Dai Inoue and Zeynettin Akkus and Timothy Kline and Weston, {Alexander D.} and Panagiotis Korfiatis and Takahashi, {Naoki M} and Erickson, {Bradley J}",
year = "2018",
month = "12",
day = "1",
doi = "10.2214/AJR.18.20331",
language = "English (US)",
volume = "211",
pages = "1184--1193",
journal = "American Journal of Roentgenology",
issn = "0361-803X",
publisher = "American Roentgen Ray Society",
number = "6",

}

TY - JOUR

T1 - What does deep learning see? Insights from a classifier trained to predict contrast enhancement phase from CT images

AU - Philbrick, Kenneth A.

AU - Yoshida, Kotaro

AU - Inoue, Dai

AU - Akkus, Zeynettin

AU - Kline, Timothy

AU - Weston, Alexander D.

AU - Korfiatis, Panagiotis

AU - Takahashi, Naoki M

AU - Erickson, Bradley J

PY - 2018/12/1

Y1 - 2018/12/1

N2 - OBJECTIVE. Deep learning has shown great promise for improving medical image classification tasks. However, knowing what aspects of an image the deep learning system uses or, in a manner of speaking, sees to make its prediction is difficult. MATERIALS AND METHODS. Within a radiologic imaging context, we investigated the utility of methods designed to identify features within images on which deep learning activates. In this study, we developed a classifier to identify contrast enhancement phase from whole-slice CT data. We then used this classifier as an easily interpretable system to explore the utility of class activation map (CAMs), gradient-weighted class activation maps (Grad-CAMs), saliency maps, guided backpropagation maps, and the saliency activation map, a novel map reported here, to identify image features the model used when performing prediction. RESULTS. All techniques identified voxels within imaging that the classifier used. SAMs had greater specificity than did guided backpropagation maps, CAMs, and Grad-CAMs at identifying voxels within imaging that the model used to perform prediction. At shallow network layers, SAMs had greater specificity than Grad-CAMs at identifying input voxels that the layers within the model used to perform prediction. CONCLUSION. As a whole, voxel-level visualizations and visualizations of the imaging features that activate shallow network layers are powerful techniques to identify features that deep learning models use when performing prediction.

AB - OBJECTIVE. Deep learning has shown great promise for improving medical image classification tasks. However, knowing what aspects of an image the deep learning system uses or, in a manner of speaking, sees to make its prediction is difficult. MATERIALS AND METHODS. Within a radiologic imaging context, we investigated the utility of methods designed to identify features within images on which deep learning activates. In this study, we developed a classifier to identify contrast enhancement phase from whole-slice CT data. We then used this classifier as an easily interpretable system to explore the utility of class activation map (CAMs), gradient-weighted class activation maps (Grad-CAMs), saliency maps, guided backpropagation maps, and the saliency activation map, a novel map reported here, to identify image features the model used when performing prediction. RESULTS. All techniques identified voxels within imaging that the classifier used. SAMs had greater specificity than did guided backpropagation maps, CAMs, and Grad-CAMs at identifying voxels within imaging that the model used to perform prediction. At shallow network layers, SAMs had greater specificity than Grad-CAMs at identifying input voxels that the layers within the model used to perform prediction. CONCLUSION. As a whole, voxel-level visualizations and visualizations of the imaging features that activate shallow network layers are powerful techniques to identify features that deep learning models use when performing prediction.

KW - Class activation map (CAM)

KW - Computer-aided diagnosis

KW - Contrast enhancement phase

KW - Convolutional neural network (CNN)

KW - CT

KW - Deep learning

KW - Gradient-weighted class activation map (Grad-CAM)

KW - Guided backpropagation

KW - Machine learning

KW - Saliency activation map

KW - Saliency map

UR - http://www.scopus.com/inward/record.url?scp=85056921390&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85056921390&partnerID=8YFLogxK

U2 - 10.2214/AJR.18.20331

DO - 10.2214/AJR.18.20331

M3 - Article

C2 - 30403527

AN - SCOPUS:85056921390

VL - 211

SP - 1184

EP - 1193

JO - American Journal of Roentgenology

JF - American Journal of Roentgenology

SN - 0361-803X

IS - 6

ER -