Update of machine learning for ultrasound diagnosis of metabolic dysfunction-associated steatotic liver disease: a bright future for deep learning

Hepatorenal index

In conventional ultrasound, the diagnosis of fatty liver is typically made by comparing the echoes of the liver and the right kidney on the same image. However, this method is affected by various factors including gain, depth, power, and patient anatomy. The hepatorenal index (HRI), a semi-quantitative biomarker of steatosis, improves upon this method by dividing the signal intensity of the liver by that of the renal cortex on the same ultrasound image (Pirmoazen et al., 2022). Higher HRI values indicate increased liver echogenicity, corresponding to higher steatosis levels. Johnson et al. (2021) demonstrated that an HRI of ≥1.4 is associated with a positive predictive value of over 95% for ≥10% steatosis. However, no corresponding guidelines currently exist. HRI is usually obtained by manually setting the region of interest (ROI) to avoid vascular and focal lesions. However, it cannot be used in the presence of ectopic kidneys or the absence of the right kidney.

Attenuation coefficient

The attenuation of sound waves in fatty liver tissue differs from that in normal liver parenchyma. The attenuation coefficient (AC) quantifies the rate of energy loss of sound waves as they propagate through tissue, which depends on the wave’s frequency and the tissue characteristics (Ozturk et al., 2023). Several algorithms are available for estimating AC, including the controlled attenuation parameter (CAP), which is the most widely studied ultrasound technique for quantifying fatty liver. CAP is measured in dB/m, with higher values indicating more severe liver steatosis. CAP is particularly sensitive in detecting fatty livers with more than 10% steatosis (Pirmoazen et al., 2020). Karlas et al. (2017) demonstrated that the diagnostic performance of CAP was reflected by an AUC value ranging from 0.823 to 0.882. However, CAP measurements are obtained without visualizing the liver, meaning they may be influenced by masses, vessels, or uneven steatosis, which can affect the accuracy of the results. Its accuracy is also diminished in the presence of obesity, ascites, and advanced fibrosis. Therefore, CAP should be combined with other ultrasound techniques to improve diagnostic reliability. In response to these limitations, researchers have developed methods for measuring liver fat content in real time on conventional ultrasound images, with the ability to accurately localize the region of interest. These methods, such as ultrasound-guided attenuation parameters (UGAP), attenuation coefficients (ATT), and attenuation imaging (ATI), utilize similar principles of attenuation measurement. For example, ATI has demonstrated greater accuracy than CAP, with an AUC ranging from 0.79 to 0.97 (Zeng et al., 2023). Despite their promising performance, UGAP, ATT, and ATI techniques are limited by factors such as operator dependence, image quality, and a lack of standardization, which can affect diagnostic accuracy.

Backscattering coefficient

The backscattering coefficient (BSC) is a quantitative value that reflects the scattering of an ultrasound pulse back to the echo probe (Wear et al., 2022). The number of scattered backward-facing ultrasound pulses increases with higher levels of liver fat because fat vesicles in liver cells enhance the scattered ultrasound signal. Lin et al. (2015) demonstrates that the backscatter coefficient (BSC) shows strong diagnostic performance in identifying hepatic steatosis, with an AUC of 0.98 for diagnosing steatosis defined by MRI-PDFF ≥5%. However, its sensitivity and specificity may vary depending on factors like BMI, and it requires specialized equipment and software for analysis.

Speed of sound

Speed of sound (SoS) can be used to characterize tissue properties based on changes in ultrasonic echo velocity in various media (Ferraioli & Monteiro, 2019). The speed of sound decreases with increasing liver fat content. Dioguardi Burgio et al. (2018) demonstrated that the novel ultrasound-based SoS provides high diagnostic performance for detecting and grading hepatic steatosis, with an AUC of 0.882 for detecting any grade of steatosis and 0.989 for moderate to severe steatosis. However, SoS is susceptible to confounding factors such as inflammation, parenchymal edema, and temperature changes.

Speckle statistics

Speckle statistics, also known as ultrasound envelope statistical parametric imaging (Park et al., 2022), analyze the speckle pattern caused by the scattering of ultrasound signals from tissue microstructures, which appears in the ultrasound image. Speckle statistics are based on the parameterization of ultrasound speckle patterns using an established statistical model that describes the scattering properties of the tissue. Acoustic structure quantification (ASQ) was introduced as a more advanced approach that builds upon the principles of speckle statistics. ASQ improves upon traditional methods by directly analyzing the backscattered ultrasound signals, offering more precise and reliable quantification of tissue composition. Lin et al. (2019) indicate that, among different scanning planes, the intercostal approach using ASQ achieves the highest AUC value of 0.92. ASQ improves precision over traditional speckle-based methods, but it may be less accessible and more dependent on equipment and scanning conditions compared to CAP.

Shear wave elastography indicators

Shear wave elastography (SWE) is an imaging technology that generates shear waves in tissues by emitting acoustic radiation pulses and converts the propagation speed of these waves into corresponding tissue hardness (Ozturk et al., 2021). Real-time, two-dimensional shear wave elastography images can be obtained through SWE, and the value of Young’s modulus (measured in kPa) of the liver can be quantitatively detected, reflecting the absolute hardness of the liver. The higher the SWE value, the harder the tissue (Taru et al., 2023). The modulus of elasticity or stiffness of diseased tissues often differs from that of normal tissues, allowing SWE to distinguish between normal and abnormal tissues. SWE is primarily used to assess liver fibrosis, an important component of MASLD and other liver diseases. It is useful in identifying patients with advanced fibrosis or cirrhosis (Castera, Friedrich-Rust & Loomba, 2019). However, for assessing steatosis, SWE showed poor correlation with steatosis grades and did not distinguish between different steatosis grades effectively (Wang et al., 2025). On the other hand, SWE could be used alongside other diagnostic tools like CAP to provide a comprehensive assessment of liver health.

Multimodal ultrasound improves the accuracy of MASLD diagnosis

The development of quantitative ultrasound holds significant promise for liver disease diagnosis, yet the establishment of standardized values remains a critical challenge, influenced by various patient and technical factors. For example, Shi et al. (2019) incorporated shear wave attenuation, shear wave absorption, elasticity, dispersion slope, and echo attenuation, achieving an impressive AUC of 0.93. While this result suggests a high degree of accuracy, the study’s reliance on a complex model may limit its clinical applicability. Similarly, Labyed & Milkowski (2020) developed the ultrasound-derived fat fraction (UDFF) by combining attenuation and backscattering coefficients, achieving an AUC range of 0.83 to 0.94 for liver fat quantification. Dillman et al. (2022) supported these findings and demonstrated similar results. While UDFF holds significant promise as a reliable, non-invasive tool for liver fat quantification, challenges related to performance variability and the need for methodological standardization must be addressed.

Moreover, quantitative ultrasound techniques have increasingly been integrated with clinical data to improve diagnostic outcomes. Notably, Newsome et al. (2020) introduced the FibroScan-AST (FAST) score, a composite score that combines vibration-controlled transient elastography (VCTE), CAP, and aspartate aminotransferase (AST), which demonstrated the best predictive properties for MASH and advanced fibrosis. In external validation, the FAST score achieved AUCs between 0.74 and 0.95, with sensitivity potentially compromised at lower AUC thresholds. More recently, the American Association for the Study of Liver Diseases (AASLD) (Rinella et al., 2023b) recommended a stepwise approach, beginning with the FIB-4 index score, followed by liver stiffness measurement (LSM) via VCTE, as the initial method for identifying high-risk MASH. While this recommendation aligns with current evidence, it highlights the challenge of developing a universally applicable screening protocol, as LSM’s effectiveness varies with liver condition and technology.

Quantitative ultrasound shows strong potential for diagnosing MASLD, but challenges remain in achieving consistency due to variability in equipment, operator skill, and patient factors. While techniques like UDFF and the FAST score show high accuracy, their performance varies, emphasizing the need for better standardization. Combining ultrasound with clinical data is promising, but creating universally applicable protocols remains challenging. Further research is needed to enhance these methods for wider clinical use.

Bayes classifier

The Bayes classifier is a probabilistic method widely used in medical diagnostics due to its ability to handle uncertainty and small datasets. In Ribeiro, Tato Marinho & Sanches (2014), a Bayes classifier was used to detect hepatic steatosis from ultrasound images, achieving 93.33% accuracy, with 94.59% sensitivity and 92.11% specificity. Similarly, Hwang & Cho (2023) applied a Bayes latent model to identify MASLD predictors, finding that ultrasound attenuation imaging (ATI) was the most effective for predicting hepatic steatosis, with an AUC of 0.923, a sensitivity of 90.2% and specificity of 85.4%. Both studies demonstrate the Bayes classifier’s potential in accurately diagnosing MASLD and hepatic steatosis non-invasively. However, the effectiveness of Bayes classifiers may be limited when there is a strong correlation or high dimensionality among features.

Support vector machine

Support vector machine (SVM) is a type of linear classifier used for supervised learning, where it separates data into different classes by finding the optimal hyperplane. Basavarajappa et al. (2021) applied SVM to six ultrasound imaging measurements, including H-mode ultrasound, and found it achieved the highest accuracy with H-mode data. Similarly, Nagy et al. (2015) used SVM to classify hepatic steatosis in 228 patients, showing that the coefficient of variation of luminance was most effective for distinguishing mild and moderate-to-severe steatosis. In summary, SVM excels in high-dimensional classification tasks but requires careful feature selection and parameter tuning to avoid overfitting or underfitting, and can be computationally intensive with large datasets. Despite these challenges, SVM remains a powerful tool in the diagnosis of MASLD due to its robustness and ability to handle complex data.

Random Forest

Random Forest (RF) is a classifier that uses multiple decision trees to improve accuracy and reduce overfitting by averaging the results from several trees. Tang et al. (2018) used RF to classify liver conditions in rats, finding that combining QUS and shear-wave elastography improved accuracy over elastography alone, though human data are still needed to confirm these results. Destrempes et al. (2022) demonstrated this view on human data. Mihai Mihailescu (2013) compared RF and SVM in assessing MASLD severity, demonstrating that RF performed better in terms of accuracy without the need for feature selection. In conclusion, RF is effective in improving classification accuracy, especially when combining multiple features. Its main advantages are high accuracy and resistance to overfitting. However, it can be computationally expensive and may require careful tuning of hyperparameters for optimal results.

Deep learning

Deep learning (DL) is a type of machine learning (ML) that builds on the development of artificial neural networks. It outperforms traditional ML, brings ML closer to Artificial Intelligence, and holds the promise of being far superior to previous techniques (Table 2). DL is not affected by image variations, does not ROIs, is not limited to expert-defined features, and can be trained on a large number of images. The algorithm uses multilayer neural networks for detection, classification, and segmentation of biomedical images. Convolutional neural networks (CNNs) are a special type of neural network in deep learning, primarily consisting of convolutional layers, pooling layers, and fully connected layers (Hosny et al., 2018) (Fig. 3). They are mainly used to process data with a grid structure, extract feature maps, and perform feature aggregation. Currently, CNNs are the most popular type of architecture in medical imaging.

Automated machine learning

Recently, a new method in machine learning, known as automated machine learning (AutoML), has been proposed. AutoML automates key aspects of the machine learning process, including data preprocessing, optimal algorithm selection, and hyperparameter tuning, significantly reducing the time required to build machine learning models (Wever et al., 2021). This automation allows for faster model development, making it more accessible to non-experts and streamlining the workflow for experienced practitioners. Tahmasebi et al. (2023) conducted a study involving 120 subjects, both with and without MASLD, using MRI-PDFF as the reference criterion. They employed AutoML Vision to develop a supervised machine learning model for assessing MASLD, achieving a sensitivity of 72.2% and a specificity of 94.6%. This study highlights AutoML’s potential in medical diagnostics, demonstrating its ability to build accurate models for complex tasks like assessing MASLD. However, despite these promising results, there is still a scarcity of relevant studies utilizing AutoML in this field, which suggests a need for further research to fully explore its capabilities and limitations.

Computer-aided design based on feature input

Computer-aided design (CAD) techniques have become increasingly integral in enhancing the diagnosis of non-alcoholic fatty liver disease (MASLD) through ultrasound imaging. The incorporation of ML in CAD systems has shown significant promise in analyzing liver ultrasound images with greater accuracy and efficiency than traditional methods. As a result, researchers have focused on developing the most effective CAD systems, exploring various classifiers to identify the best-performing models for diagnosing MASLD.

Early studies, such as those by Acharya et al. (2012), employed the Symtosis CAD method to extract features from liver ultrasound images, combining texture analysis, higher-order spectra, and wavelet transforms. This technique, paired with a decision tree (DT) classifier, accurately distinguished between normal and fatty liver images. Similarly, Subramanya et al. (2014) used a SVM-based CAD system to diagnose fatty liver, finding that Laws ratio features yielded the best classification results.

As technology advanced, CAD systems evolved, incorporating more sophisticated algorithms and a broader range of features. Saba et al. (2016) applied Symtosis to extract additional features, including Haralick, Gabor, and Fourier transforms, improving the system’s ability to capture subtle liver texture changes. Their study showed that a back propagation neural network (BPNN) classifier outperformed traditional DT classifiers. Acharya et al. (2016) introduced a curvelet-wave transform method with a probabilistic neural network (PNN), which showed high classification accuracy using only six features. These advancements marked a shift toward more efficient and robust models, reducing the complexity of the feature set required for accurate classification.

Further improvements were made in more recent studies. Kuppili et al. (2017) and Biswas et al. (2018) explored the combination of Symtosis with extreme learning machines (ELM) and CNNs, both of which led to faster and more accurate liver image classification. These models capitalized on the ability of CNNs to automatically learn and refine features from raw data, removing the need for extensive manual feature extraction.

While CAD systems using various classifiers have been developed for diagnosing hepatic steatosis, several challenges remain. Factors such as image quality, feature selection, and the complexity of classification algorithms can all impact the performance of these systems. The application of machine learning techniques like SVM, CNN, and PNN has significantly improved liver image classification accuracy. The future of ultrasonic CAD holds great promise, particularly in enhancing the early detection and management of MASLD, offering more reliable and efficient diagnostic tools.