Classification for Breast Ultrasound Using Convolutional Neural Network with Multiple Time-Domain Feature Maps

Geuntae Im,Heung Jae Lee,Jongsoo Lee,Ki-Baek Lee,Hyungsuk Kim,Hakjoon Lee,Juyoung Park

doi:10.3390/app112110216

Abstract

Ultrasound (US) imaging is widely utilized as a diagnostic screening method, and deep learning has recently drawn attention for the analysis of US images for the pathological status of tissues. While low image quality and poor reproducibility are the common obstacles in US analysis, the small size of the dataset is a new limitation for deep learning due to lack of generalization. In this work, a convolutional neural network (CNN) using multiple feature maps, such as entropy and phase images, as well as a B-mode image, was proposed to classify breast US images. Although B-mode images contain both anatomical and textual information, traditional CNNs experience difficulties in abstracting features automatically, especially with small datasets. For the proposed CNN framework, two distinct feature maps were obtained from a B-mode image and utilized as new inputs for training the CNN. These feature maps can also be made from the evaluation data and applied to the CNN separately for the final classification decision. The experimental results with 780 breast US images in three categories of benign, malignant, and normal, showed that the proposed CNN framework using multiple feature maps exhibited better performances than the traditional CNN with B-mode only for most deep network models.

Highlights

Among medical imaging modalities, ultrasound (US) is one of the most commonly utilized in clinical screening and diagnostic applications due to its safety by utilization of non-ionizing radiation, portability, cost effectiveness, and real-time data acquisition and display
Since the main purpose of this work is the benefits of multiple time-domain feature maps in CNNbased deep learning, no artificial manipulation was applied to compensate the dataset imbalance and classification performances were compared under the same conditions for the traditional convolutional neural network (CNN) and the proposed featurechannel convolutional neural network (FC-CNN) methods
We used four different combining algorithms in the decision layer to produce the final prediction, the results of the proposed FC-CNN method shown in Table 2 were obtained using the simple voting algorithm, which achieved the best performance

Summary

Introduction

Ultrasound (US) is one of the most commonly utilized in clinical screening and diagnostic applications due to its safety by utilization of non-ionizing radiation, portability, cost effectiveness, and real-time data acquisition and display. Despite these advantages, US imaging has limitations, such as relatively low imaging contrast and degradation of quality caused by noise and speckles, high image variability due to the operator-dependent hand-held nature in the data acquisition process, and poor image reproducibility across different manufacturers’ US imaging systems. Since AlexNet [2], which is a convolutional neural network (CNN)

Objectives

Methods

Results

Discussion

Conclusion