Steatosis Quantification on Ultrasound Images by a Deep Learning Algorithm on Patients Undergoing Weight Changes

Adam P. Harrison, Bowen Li, Tse-Hwa Hsu, Cheng‐Jen Chen, Wanting Yu, Jennifer Tai, Le Lü, Dar‐In TaiDiagnostics 2023 · DOI

TL;DR

This paper introduces a deep learning (DL) algorithm for quantifying liver steatosis from standard ultrasound images, addressing the subjectivity of current methods. The algorithm demonstrates a strong correlation with body weight changes and gender, suggesting its potential as an objective, accessible tool for monitoring liver health.

What Problem Does This Paper Address?

Non-alcoholic fatty liver disease (NAFLD) is a growing global health concern, with liver steatosis being a primary indicator. While liver histology is the gold standard for diagnosis, it's invasive. MRI-PDFF is a non-invasive alternative but is expensive. Standard 2D ultrasound (US) is widely used but its assessment of steatosis is highly subjective. Existing quantitative US (QUS) methods often require specialized equipment and are not easily interchangeable. This paper aims to overcome these limitations by developing and validating a DL algorithm that can objectively quantify liver steatosis using readily available 2D US images, specifically in the context of patients undergoing weight changes.

Methods

The core of this research is a deep learning algorithm, previously developed and trained on a large dataset of ultrasound images. This algorithm takes standard 2D B-mode US images as input and outputs a continuous steatosis severity score (ranging from 0 to 1). For this study, the researchers focused on validating this algorithm in a cohort of patients who experienced significant weight changes. They classified ultrasound images into four scanning views (G1: left lobe, G2: right intercostal, G3: liver/kidney contrast, G4: subcostal) and used a protocol to select the best view (preferentially G2, then G1 or G4) for analysis. The algorithm was applied to 3-5 images per view to derive a mean steatosis score for each study. These scores were then correlated with patient body weight changes over time. Statistical analysis included repeated measures correlation to account for longitudinal data and paired t-tests to compare steatosis scores between high and low body weight timepoints.

Datasets

The study utilized a retrospective cohort of 74 patients from the Chang Gung Memorial Hospital, with a total of 2393 images from 239 studies collected between 2013 and 2021. These patients were selected based on having experienced a weight change of ≥5%. The DL algorithm itself was previously trained on a much larger dataset of 2899 patients and 200,654 images acquired across 13 different scanners.

Key Insights and Ideas

  • Bridging the Subjectivity Gap: The DL algorithm offers a crucial step towards objective quantification of liver steatosis from standard US, moving beyond subjective visual interpretation.
  • Weight as a Proxy for Steatosis: The strong correlation between body weight changes and steatosis scores validates the algorithm's ability to reflect physiological changes in liver fat content. This is particularly important for patients undergoing weight management.
  • Optimal Imaging Protocol: The study identified that 3-5 images from the right intercostal view (G2) provide the best scanning conditions for reliable steatosis quantification.
  • Gender Differences in Steatosis: The algorithm revealed that men tend to have a higher liver steatosis/BMI ratio than women, a finding that became more pronounced at higher body weights. This highlights potential sex-specific differences in fat accumulation and NAFLD progression.
  • Accessibility and Flexibility: Unlike specialized QUS techniques, this DL algorithm can be applied to images from standard US machines and does not require defining a specific region of interest, making it more practical for widespread clinical use.

Key Results

  • The DL algorithm showed high correlation between different scanning views (G1 vs. G2, G2 vs. G4) with R² > 0.86 and p < 0.001.
  • A significant correlation was found between body weight and steatosis scores (R² = 0.62, p < 0.001).
  • Significant differences in steatosis scores were observed between the highest and lowest body weight timepoints (p < 0.001).
  • Men exhibited a higher liver steatosis/BMI ratio than women (p = 0.026).
  • The optimal number of images for repeatable quantification was found to be 3-5, particularly from the right intercostal view (G2).

Limitations and Future Work

The authors acknowledge several limitations:

  • Retrospective Design: The study was retrospective, and histology or MRI-PDFF was not used for direct validation of the DL algorithm's output in this cohort.
  • Scanner Variability: While the algorithm was trained on multiple scanners, the use of different machines across studies could introduce some variability. However, previous work indicated high agreement between scanners.
  • Weight as the Sole Metric: Body weight was used as the primary quantitative standard for correlating with steatosis. The study noted that weight loss is not the only factor influencing steatosis, citing an example where the lowest steatosis score occurred at the beginning of weight change, not necessarily at the lowest weight.
  • Small Case Number: The relatively small sample size of 74 patients is a limitation.

Future work should focus on prospective external validation with larger cohorts, potentially including histology or MRI-PDFF as gold standards. Further investigation into the complex interplay of factors beyond weight (e.g., diet, exercise, hormonal influences) and their impact on steatosis quantification is also warranted. The authors suggest establishing a standardized protocol for acquiring 3-5 good-quality liver images from the right intercostal view (or alternatives if unavailable) for future retrospective quantification.

Summary updated

Abstract

The best scanning conditions are 3-5 images from the right intercostal view. The algorithm objectively quantified liver steatosis, which correlated with body weight changes and gender.
BibTeX
@inproceedings{harrison2023steatosis,
  author = {Adam P. Harrison and Bowen Li and Tse-Hwa Hsu and Cheng‐Jen Chen and Wanting Yu and Jennifer Tai and Le Lü and Dar‐In Tai},
  title = {{Steatosis Quantification on Ultrasound Images by a Deep Learning Algorithm on Patients Undergoing Weight Changes}},
  booktitle = {Diagnostics},
  year = {2023},
  doi = {10.3390/diagnostics13203225},
}
Figure