Deep learning-based ultrasound models in the evaluation of magnetic resonance imaging-proton density fat fraction

Abstract

Abstract: Objective The use of Proton Density Fat Fraction (PDFF) in the diagnosis of fatty liver has been limited by its high cost and limited availability. This study aimed to develop a liver ultrasound model using deep learning for the evaluation of fatty liver severity and steatosis assessment based on PDFF. Methods Patients whose liver ultrasound and MRI-PDFF examinations were completed within one week in our hospital were included in the study. Two models were developed: 1) a regression model: quantifying the degree of liver steatosis, with continuous variable PDFF values as predictive data labels; 2) a classification model: classifiers were developed based on three threshold values of PDFF (5%, 10%, and 25%) to qualitatively diagnose and assess the severity of fatty liver. Results Predictions from the regression model based on three views (subcostal longitudinal, second liver portal, and hepatorenal interface) were correlated with actual PDFF values using Spearman correlation analysis. The results showed a close correlation between the predicted values and actual values (r values of 0.860, 0.922, and 0.903, respectively). Moreover, the average value of the three views correlated with the actual PDFF value with an r value of 0.961. The classification model based on the three views for binary classification using a PDFF 5% as the threshold showed significant performance (AUCs of 0.938, 0.946, and 0.959, respectively). The average value of the three views achieved an AUC of 0.970. As PDFF 10% as the threshold, the performance was also significant (AUCs of 0.974, 0.991, and 0.997), with the average achieving 0.998. As PDFF 25% as the threshold, the three AUCs remained significant (0.904, 0.958, and 0.936), with the average AUC being 0.992. Finally, to visualize the deep learning model in liver ultrasound prediction, we used Gradient-weighted Class Activation Mapping for visual interpretation of the convolutional neural network. Conclusion This study proposes deep learning-based ultrasound image classification and quantification models for fatty liver. These models have the potential to improve efficiency and reduce costs in the diagnosis and treatment of fatty liver, and shows significant clinical application prospects.

Key words: Nonalcoholic fatty liver disease, Chemical-shift-encoded magnetic resonance imaging (CSE-MRI), Proton density fat fraction (PDFF), Ultrasound, Deep learning, Transfer learning

GE Hai-long, CHAO Chen, SHI Xing-song, BAI Cui-wei, WANG Yu, YIN Qi. Deep learning-based ultrasound models in the evaluation of magnetic resonance imaging-proton density fat fraction[J]. Chinese Hepatolgy, 2025, 30(10): 1393-1397.

References

[1] Zhu J Z, Dai Y N, Wang Y M, et al. Prevalence of nonalcoholic fatty liver disease and economy[J]. Dig Dis Sci,2015,60(11):3194-3202.
[2] 杨晨,余上海,徐飞鹏,等.磁共振质子密度脂肪分数(MRI-PDFF)在肝脂肪定量中的应用[J].临床肝胆病杂志,2024,40(3):600-605.
[3] Tamaki N, Ajmera V, Loomba R. Non-invasive methods for imaging hepatic steatosis and their clinical importance in NAFLD[J]. Nat Rev Endocrinol,2022,18(1):55-66.
[4] Nogami A, Yoneda M, Iwaki M, et al. Non-invasive imaging biomarkers for liver steatosis in non-alcoholic fatty liver disease: present and future[J]. Clin Mol Hepatol,2023,29:S123-S135.
[5] Yin M, Liu L, Gao J, et al. Deep learning for pancreatic diseases based on endoscopic ultrasound: a systematic review[J]. Int J Med Inform,2023,174:105044.
[6] 中华医学会肝病学分会脂肪肝和酒精性肝病学组,中国医师协会脂肪性肝病专家委员会.非酒精性脂肪性肝病防治指南(2018更新版)[J].中华肝脏病杂志,2018,26(3):195-203.
[7] Hong C W, Marsh A, Wolfson T, et al. Reader agreement and accuracy of ultrasound features for hepatic steatosis[J]. Abdom Radiol (NY),2019,44(1):54-64.
[8] Castera L, Friedrich-Rust M, Loomba R. Noninvasive assessment of liver disease in patients with nonalcoholic fatty liver disease[J]. Gastroenterology,2019,156(5):1264-1281.
[9] Permutt Z, Le T A, Peterson M R, Seki E, et al. Correlation between liver histology and novel magnetic resonance imaging in adult patients with non-alcoholic fatty liver disease-MRI accurately quantifies hepatic steatosis in NAFLD[J]. Aliment Pharmacol Ther,2012,36(1):22-29.
[10] Wang Y, Hong Y, Wang Y, et al. Automated multimodal machine learning for esophageal variceal bleeding prediction based on endoscopy and structured data[J]. J Digit Imaging,2023,36(1):326-338.
[11] 高静雯,林嘉希,刘璐,等.基于深度学习二次预训练建立Barrett食管内镜图片分类模型的研究[J].中国数字医学,2022,17(10):54-58.
[12] Tapper E B, Lok A S F. Use of liver imaging and biopsy in clinical practice[J]. N Engl J Med,2017,377(23):2296-2297.
[13] Vianna P, Calce S I, Boustros P, et al. Comparison of radiologists and deep learning for US grading of hepatic steatosis[J]. Radiology,2023,309(1):e230659.
[14] He Y, Lin J, Zhu S, et al. Deep learning in the radiologic diagnosis of osteoporosis: a literature review[J]. J Int Med Res,2024,52(4):3000605241244754.
[15] Byra M, Styczynski G, Szmigielski C, et al. Transfer learning with deep convolutional neural network for liver steatosis assessment in ultrasound images[J]. Int J Comput Assist Radiol Surg,2018,13(12):1895-1903.
[16] Reddy D, Bharath R, Rajalakshmi P. A novel computer-aided diagnosis framework using deep learning for classification of fatty liver disease in ultrasound imaging[C]. IEEE,2018,pp,1-5.
[17] Byra M, Han A, Boehringer A S, Zhang Y N, et al. Liver fat assessment in multiview sonography using transfer learning with convolutional neural networks[J]. J Ultrasound Med,2022,41(1):175-184.