Development and Preliminary Evaluation of an EfficientNet-Based Deep Learning System for Ultrasound Assessment of Neck Disorders: A Single-Center Study

TL;DR

This study develops a deep learning system using ultrasound images to help diagnose neck disorders, showing promising accuracy in assessing specific regions of the neck.

Contribution

The novel contribution is the application of an EfficientNet-based model with transfer learning to ultrasound data for diagnosing neck disorders.

Findings

The model achieved 76% and 81% weighted average F1-scores for deep fascia texture and myofascial adhesion in lower cervical regions.

Macro averages of 74% and 78% indicate consistent class-wise accuracy for these regions.

The system shows robust classification performance for neck disorder assessment.

Abstract

Background/Objectives: Neck disorders encompass a range of discomforts impacting a person’s quality of life. Traditional diagnostic methods, such as physical tests and imaging techniques, rely heavily on clinician expertise, leading to potential variability in assessments. While ultrasound imaging is commonly used, the application of machine learning models to assess neck disorders, particularly fascial abnormalities, remains limited. This study seeks to fill this gap by developing a machine learning model using ultrasound images to provide accurate and efficient support for diagnosing neck disorders. Methods: Due to limited availability of labeled ultrasound data for neck disorders, developing robust and generalizable models remains a challenge. In this study, a neck disorder assessment system was developed using ultrasound images collected from 184 patients by employing various machine learning algorithms. To address data scarcity and improve model generalizability, an approach utilizing EfficientNet with transfer learning was introduced and thoroughly assessed using the trained model on a completely clean test dataset, ensuring the robustness of the solution. The model was trained using 5-fold cross-validation with the respective weight of each class and AdamW as the optimizer. Results: The results showed promising performance, with the deep fascia fuzzy texture and deep fascia and myofascial adhesion at lower cervical regions demonstrating the highest weighted average F1-scores of 76% and 81%, respectively. The macro averages reflected similar performance, at 74% and 78%, respectively, indicating consistent class-wise accuracy for these regions. Conclusions: The proposed model demonstrated robust classification performance for neck disorder assessment, particularly in evaluating the lower cervical region. This approach has the potential to support clinical decision-making by providing consistent, efficient, and accurate diagnostic assistance. Further refinement and validation across diverse clinical settings will be critical to enhance its real-world applicability.

Figures