A02-KB006 Computer-Aided Diagnosis System Using Image Features and Metabolic Function for Lesion on Breast MRI

From Multidisciplinary Computional Anatomy
Jump to: navigation, search

Member

  • Primary Investigator
    Ryohei Nakayama (Ritsumeikan University, Associate Professor)
  • Co-Investigator
    Emi Honda (Japanese Red Cross Medical Center)
    Kiyoshi Namba (Hokuto Hospital and Clinic)

Overview

Our purpose in this project was to develop a computer-aided diagnosis (CAD) scheme for distinguishing between benign and malignant breast masses using both image features and metabolic function for lesion on breast magnetic resonance (MR) imaging. Our CAD scheme evaluated the likelihood of malignancy by inputting dynamic contrast material-enhanced images, T2-weighted images, and diffusion weighted images for the same patient into a Convolutional Neural Network. With the CAD scheme, the classification accuracy, sensitivity, and specificity were shown to be 93.3% (84 of 90), 95.2% (59 of 62), and 89.3% (25 of 28), respectively. The positive and negative predictive values were 95.2% (59 of 62) and 89.3% (25 of 28), respectively. In the observer study, the areas under the ROC curves (AUCs) for all radiologists were improved by use of the CAD scheme. The average AUC increased from 0.618 without to 0.775 with the CAD scheme (P = .017). Our CAD scheme was shown to have high classification accuracy and be useful in the differential diagnosis of masses on MRI images.

Project Design

A02-KB006-E.png