Background: Hepatic Steatosis is one of the most prevalent
liver disorders globally. Ultrasound imaging is widely used as the primary
screening tool for Hepatic Steatosis. However, its diagnostic performance can
vary significantly depending on the operator’s skill and the quality of the
equipment. Recent advances in deep learning have brought new opportunities to
medical imaging, providing automated, consistent, and quantitative assessments
that reduce dependency on operator expertise.
Objectives: This study aims to develop a deep learning
(DL)-based framework that enhances the detection and grading of Hepatic
Steatosis from ultrasound images. The key goal is to achieve accuracy levels
comparable to experienced radiologists while maintaining interpretability and
efficiency for real-time use in clinical practice.
Methods: B-mode ultrasound images and cine clips were
collected from patients, covering multiple liver views to capture diverse
anatomical perspectives. Alongside imaging data, patient metadata such as age,
body mass index (BMI), and comorbid conditions were also recorded to enrich the
dataset. The proposed system employs a multi-view ultrasound preprocessing
approach, followed by transfer learning to leverage existing feature
representations. Attention-driven convolutional neural networks (CNNs) are then
used to capture fine details across image regions. To ensure clinical
usability, explainability modules are integrated, allowing transparent
interpretation of model predictions.
Findings: Experimental evaluation demonstrated that the
framework outperformed traditional single-view methods, offering improved
sensitivity and specificity in detecting hepatic Steatosis. The performance was
closely aligned with radiologist-level assessments. Furthermore, the system
showed low latency, highlighting its suitability for near-real-time diagnostic
applications.
Conclusion: Unlike conventional models that rely on a single
static image, this study introduces a multi-view fusion strategy enhanced with
attention mechanisms and explainability tools. This combination not only
strengthens predictive accuracy but also ensures transparency and
trustworthiness, critical factors for adoption in clinical settings. Despite
the promising performance, challenges such as data variability, subtle
early-stage disease patterns, and model interpretability remain. Addressing these
limitations through larger, diverse datasets and explainable AI approaches will
be essential for translating these models into clinical practice.
Author(s) Details
A. Sahaya Mercy
Department of Computer Science, St. Joseph’s College (Autonomous), Tiruchirappalli-2,
Affiliated to Bharathidasan University, Tamil Nadu, India.
G. Arockia Sahaya
Sheela
Department of Computer Science, St. Joseph’s College (Autonomous),
Tiruchirappalli-2, Affiliated to Bharathidasan University, Tamil Nadu, India.
Please see the book here :- https://doi.org/10.9734/bpi/mcsru/v8/6581
