Automated scan quality evaluation for DDH using transfer learning: Development of a novel ensemble system

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Pediatric Orthopedics

Background:

• Developmental Dysplasia of the Hip (DDH) is a common infant condition requiring early diagnosis for effective non-invasive treatment.
• Accurate diagnosis via ultrasound (US) relies on precise identification of anatomical landmarks.
• Variability in landmark identification by observers can impact diagnostic accuracy.

Purpose of the Study:

• To develop an automated system for assessing pelvic US image quality.
• To evaluate the quality of five anatomical landmarks using transfer learning models.
• To address intra-observer and inter-observer variability in DDH diagnosis.

Main Methods:

• Utilized US images from 1,891 subjects across two hospitals in Korea.
• Developed an ensemble system with transfer learning models for automated scan quality evaluation.
• Employed Gradient-weighted class activation mapping for model verification and proposed an Alternative Sequence Method (ASM) for real-time application.

Main Results:

• Selected models achieved kappa values of 0.6 or higher, indicating substantial agreement.
• The AUC score for classifying standard images was 0.89.
• The proposed ASM demonstrated improved scan quality assessment efficiency, with classification times of 0.27, 0.22, and 0.20 seconds per image for ASM-1, ASM-2, and ASM-3, respectively, compared to 0.35 seconds for the full sequence method.

Conclusions:

• The automated system effectively assesses pelvic US image quality for DDH diagnosis.
• Transfer learning models show promise in reducing observer variability and improving diagnostic consistency.
• The ASM offers a faster and efficient alternative for real-time scan quality assessment in clinical settings.