Deep Neural Architectures for Contrast Enhanced Ultrasound (CEUS) Focal Liver Lesions Automated Diagnosis

Abstract

Computer vision, biomedical image processing and deep learning are related fields with a tremendous impact on the interpretation of medical images today. Among biomedical image sensing modalities, ultrasound (US) is one of the most widely used in practice, since it is noninvasive, accessible, and cheap. Its main drawback, compared to other imaging modalities, like computed tomography (CT) or magnetic resonance imaging (MRI), consists of the increased dependence on the human operator. One important step toward reducing this dependence is the implementation of a computer-aided diagnosis (CAD) system for US imaging. The aim of the paper is to examine the application of contrast enhanced ultrasound imaging (CEUS) to the problem of automated focal liver lesion (FLL) diagnosis using deep neural networks (DNN). Custom DNN designs are compared with state-of-the-art architectures, either pre-trained or trained from scratch. Our work improves on and broadens previous work in the field in several aspects, e.g., a novel leave-one-patient-out evaluation procedure, which further enabled us to formulate a hard-voting classification scheme. We show the effectiveness of our models, i.e., 88% accuracy reported against a higher number of liver lesion types: hepatocellular carcinomas (HCC), hypervascular metastases (HYPERM), hypovascular metastases (HYPOM), hemangiomas (HEM), and focal nodular hyperplasia (FNH).

Interpretable Machine Learning for Characterization of Focal Liver Lesions by Contrast-Enhanced Ultrasound

“…Comparing the obtained performance with similar methods for computer-aided diagnosis of FLLs, accuracies ranging from 85.8% to 91.8% have been reported [ 21 ]–[ 26 ], [ 28 ]. However, a fair comparison is difficult, as other methods might have different objectives [ 24 ], [ 26 ], [ 28 ], and they have been tested on a different population, sometimes using larger training datasets [ 22 ], [ 24 ].…”

“…However, a fair comparison is difficult, as other methods might have different objectives [ 24 ], [ 26 ], [ 28 ], and they have been tested on a different population, sometimes using larger training datasets [ 22 ], [ 24 ]. In addition, they generally require extra dedicated procedures for compensation of in-plane motion, manual selection of out-of-plane frames, and selection of a parenchyma ROI [ 21 ]–[ 26 ], [ 28 ].…”

“…A number of methods have been proposed to quantitatively analyze CEUS images for computer-aided FLLs characterization [ 15 ], [ 20 ]–[ 28 ]. Initial attempts mainly focused on assessing the temporal characteristics US contrast enhancement by analysis of TICs [ 15 ], [ 20 ], [ 22 ], [ 23 ].…”

“…Recently, a method based on convolutional neural networks was also proposed to achieve end-to-end lesion classification, avoiding the need for hand-engineered feature extraction [ 28 ]. The method, however, still requires the manual selection of the lesion ROI by an expert radiologist, based on which 50 ROIs from each vascular phase (arterial, portal, and late) are then manually selected.…”

“…Some methods additionally require long acquisition times up to 30 min [ 21 ], [ 24 ], [ 25 ] to observe the postvascular phase and the injection of a second bolus followed by a high-pressure US pulse to sample the replenishment curve by inflowing microbubbles [ 24 ]. More complex models can achieve end-to-end classification, avoiding the need for all preprocessing and feature extraction steps, but often require large patient dataset due to the large number of trainable model parameters [ 26 ], [ 28 ] and they are generally more difficult to interpret [ 23 ], [ 26 ], [ 28 ].…”