Improving B-mode ultrasound diagnostic performance for focal liver lesions using deep learning: A multicentre study - PubMed (original) (raw)

Comparative Study

doi: 10.1016/j.ebiom.2020.102777. Epub 2020 Apr 28.

Jingwei Wei 2, Xiaohan Hao 3, Dexing Kong 4, Xiaoling Yu 1, Tianan Jiang 5, Junqing Xi 1, Wenjia Cai 1, Yanchun Luo 1, Xiang Jing 6, Yilin Yang 7, Zhigang Cheng 1, Jinyu Wu 8, Huiping Zhang 9, Jintang Liao 10, Pei Zhou 11, Yu Song 12, Yao Zhang 13, Zhiyu Han 1, Wen Cheng 14, Lina Tang 15, Fangyi Liu 1, Jianping Dou 1, Rongqin Zheng 16, Jie Yu 17, Jie Tian 18, Ping Liang 19

Affiliations

Comparative Study

Improving B-mode ultrasound diagnostic performance for focal liver lesions using deep learning: A multicentre study

Qi Yang et al. EBioMedicine. 2020 Jun.

Abstract

Background: The diagnosis performance of B-mode ultrasound (US) for focal liver lesions (FLLs) is relatively limited. We aimed to develop a deep convolutional neural network of US (DCNN-US) for aiding radiologists in classification of malignant from benign FLLs.

Materials and methods: This study was conducted in 13 hospitals and finally 2143 patients with 24,343 US images were enrolled. Patients who had non-cystic FLLs with pathological results were enrolled. The FLLs from 11 hospitals were randomly divided into training and internal validations (IV) cohorts with a 4:1 ratio for developing and evaluating DCNN-US. Diagnostic performance of the model was verified using external validation (EV) cohort from another two hospitals. The diagnosis value of DCNN-US was compared with that of contrast enhanced computed tomography (CT)/magnetic resonance image (MRI) and 236 radiologists, respectively.

Findings: The AUC of ModelLBC for FLLs was 0.924 (95% CI: 0.889-0.959) in the EV cohort. The diagnostic sensitivity and specificity of ModelLBC were superior to 15-year skilled radiologists (86.5% vs 76.1%, p = 0.0084 and 85.5% vs 76.9%, p = 0.0051, respectively). Accuracy of ModelLBC was comparable to that of contrast enhanced CT (both 84.7%) but inferior to contrast enhanced MRI (87.9%) for lesions detected by US.

Interpretation: DCNN-US with high sensitivity and specificity in diagnosing FLLs shows its potential to assist less-experienced radiologists in improving their performance and lowering their dependence on sectional imaging in liver cancer diagnosis.

Keywords: Convolutional neural network; Diagnosis; Focal liver lesions; Ultrasound.

Copyright © 2020 The Authors. Published by Elsevier B.V. All rights reserved.

PubMed Disclaimer

Conflict of interest statement

Declaration of Competing Interest The authors report no conflicts of interest.

Figures

Fig. 1

Fig. 1

Flowchart: development and validation of DCNN-US for diagnosis of focal liver lesions. US=Ultrasound, LB=Liver background, HBV=Heaptitis B virus, HCV=Heaptitis C virus, ROI=Region of interest, BN=Batch normalization, CT=Computed tomography, MRI=Magnetic resonance image, DCNN-US=Deep convolutional neural network of ultrasound.

Fig. 2

Fig. 2

Model robustness analysis. (a-c) represents 1st fold in training, internal validation and external validation cohort, respective. The green, blue, red line represents the ROC curve of ModelL, ModelLB, ModelLBC, respectively. ROC=Receiver operating characteristic, ModelL=Modellesion, ModelLB=Modellesion+background, ModelLBC=Modellesion+background+clinic.

Fig. 3

Fig. 3

Individualized predictive graphical presentation for clinical use. (a) Nomogram for the DCNN-US model. By integrating the predicted score, clinical factors including gender, age, tumour history, hepatitis history, hypoechoic halo, and vascularity of lesion and the final diagnostic outcome are presented on the bottom line of the nomogram along with malignancy probability. Gender, 0: female, 1: male; Tumour history, 0: no, 1: yes; Hepatitis history, 0: none, 1: HBV, 2: HCV, 3: HBV+HCV; Hypoechoic halo, 0: absent, 1 present; Vascularity, 0: absent, 1: present. (b) Calibration curve of the training cohort's nomogram. (c) Calibration curve of the internal validation cohort's nomogram. (d) Calibration curve of the external validation cohort's nomogram. Calibration curves indicate the consistency between histological diagnosis and predicted malignancy scores. The blue dotted line represents a perfect prediction by an ideal model. The pink solid line represents the nomogram's performance. A closer distance of the pink line to the blue line represents a better prediction. The p value of the Hosmer–Lemeshow test was greater than 0.05 for both the training and internal validation cohorts, showing good calibration between predictive outcome and histological diagnosis. (e) Decision curve analysis. The y-axis represents net benefit. The yellow and green lines measure the benefit obtained from ModelLB and ModelLBC, respectively. The blue and black lines measure the benefit of using the “all malignant FLLs” and “all benign FLLs” strategies, respectively. DCNN-US=Deep convolutional neural network of ultrasound, FLLs=Focal liver lesions, ModelLB=Modellesion+background, ModelLBC=Modellesion+background+clinic.

Fig. 4

Fig. 4

Classification performance of the DCNN-US model and radiologists on focal liver lesions. (a-c) represents accuracy, sensitivity and specificity comparison between 15-year skilled radiologists and DCNN-US model, respectively. All p values were performed by non-parametric test. (d) Bar graph shows the diagnostic performance of the DCNN-US model, contrast enhanced CT and MRI. CT=Computed tomography, MRI=Magnetic resonance imaging, DCNN-US=Deep convolutional neural network of ultrasound.

Fig. 5

Fig. 5

Attention maps of model on the benign and malignant lesions. Colours from warm to cold represent the degree of pixels’ contribution to FLL diagnosis. Red indicates the areas that contributed most, and blue areas contributed least. The number in the picture indicates the malignancy probability predicted by the model. (a) Hepatocellular carcinoma with different ultrasound appearance (Left MP: 0.9988, Right MP: 0.9989). (b) Metastatic cancer with different ultrasound appearance (Left MP: 0.9497, Right MP: 0.9997). (c) Hepatocellular carcinoma (MP: 0.9981) with similar ultrasound appearance to hepatic adenoma (MP: 0.2099). (d) Cholangiocellular carcinoma (MP: 0.8686) with similar ultrasound appearance to focal nodular hyperplasia (MP: 0.1393). FLL=Focal liver lesion, MP=malignancy probability predicted by the model.

Comment in

References

    1. Bray Freddie, Ferlay Jacques, Soerjomataram Isabelle. Global Cancer Statistics 2018: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA Cancer J Clin. 2018;0:1–31. - PubMed
    1. Amarapurkar Deepak, Han Kwang-Hyub, Chan Henry Lik-Yuen. Application of surveillance programs for hepatocellular carcinoma in the Asia–Pacific Region. J Gastroenterol Hepatol. 2009;24:955–961. - PubMed
    1. Marrero J.A., Ahn J., Rajender R.K. ACG clinical guideline: the diagnosis and management of focal liver lesions. Am J Gastroenterol. 2014;109(9):1328–1347. - PubMed
    1. Heimbach Julie K., Kulik Laura M., Finn Richard S. AASLD Guidelines for the treatment of hepatocellular carcinoma. Hepatology. 2018;67(1):358–380. - PubMed
    1. Tchelepi H., Ralls P.W. Ultrasound of focal liver masses. Ultrasound Q. 2004;20(4):155–169. - PubMed

Publication types

MeSH terms

Substances

LinkOut - more resources