2025 ISAKOS Biennial Congress Paper
Deep Learning-Based Ultrasonographic Imaging Diagnosis of Chronic Lateral Ankle Instability
Tetsuya Yamamoto, MD, PhD, Kobe, Hyogo JAPAN
Noriyuki Kanzaki, MD, PhD, Kobe, Hyogo JAPAN
Yuta Nakanishi, MD, PhD, Kobe, Hyogo JAPAN
Kyohei Nishida , MD, PhD, Kobe, Hyogo JAPAN
Kanto Nagai, MD, PhD, Kobe, Hyogo JAPAN
Yuichi Hoshino, MD, PhD, Kobe, Hyogo JAPAN
Takehiko Matsushita, MD, PhD, Kobe, Hyogo JAPAN
Ryosuke Kuroda, MD, PhD, Kobe, Hyogo JAPAN
Kobe University Graduate School of Medicine, Kobe, Hyogo, JAPAN
FDA Status Not Applicable
Summary
The deep learning can significantly improve the accuracy and reliability of ultrasonographic diagnosis of chronic lateral ankle instability.
Abstract
Background
Chronic lateral ankle instability (CLAI) affects 10-30% of individuals who suffer from repeated ankle sprains. Accurate diagnosis is essential for effective treatment and management. Ultrasonography has proven to be a valuable tool for assessing the ankle and foot, particularly in evaluating the anterior talofibular ligament (ATFL) under anterior drawer stress. However, diagnosing CLAI with ultrasonography requires significant proficiency and expertise. The integration of deep learning techniques with ultrasonographic imaging could offer a promising approach to improving diagnostic accuracy and objectivity.
Objective
The purpose of this study was to evaluate the usefulness of deep learning models in the ultrasonographic diagnosis of CLAI.
Study Design
The study involved 20 feet from patients diagnosed with CLAI based on stress X-ray and physical examination at our hospital. An anterior drawer stress test was conducted on these patients, and 300 still images were extracted at maximum anterior stress from the ultrasonographic videos. A control group consisting of 20 healthy feet with no history of ankle sprain underwent the same imaging procedure, yielding another set of 300 still images. The images were then divided into training and test datasets, with 80% allocated for training and 20% for testing. Three types of network models were trained using transfer learning via the Deep Learning Toolbox (MathWorks). The performance of these models was evaluated using the test data, focusing on metrics such as accuracy, precision, recall, specificity, and F-measure. Additionally, the area under the curve (AUC) from the receiver operating characteristic (ROC) curve was calculated to assess overall diagnostic accuracy.
Results
Among the trained models, the most accurate achieved an accuracy rate of 0.84, a precision rate of 0.79, a recall rate of 0.88, a specificity of 0.80, and an F-measure of 0.83. The area under the curve (AUC) from the ROC curve was 0.94, reflecting a high level of diagnostic accuracy.
Conclusion
Deep learning can be effectively utilized in the ultrasonographic diagnosis of CLAI with a high degree of diagnostic accuracy. This innovative approach shows promise for enhancing the objectivity and reliability of ultrasonographic diagnoses of CLAI. Implementing such deep learning systems in clinical practice could assist practitioners by providing precise and consistent diagnostic insights, potentially leading to improved patient outcomes.