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

## TL;DR

This paper introduces an automated system using transfer learning to evaluate the quality of ultrasound images for hip disorders in infants, aiming to improve early diagnosis accuracy and reduce variability.

## Contribution

A novel ensemble system using transfer learning and an alternative sequence method for real-time scan quality evaluation in DDH diagnosis is proposed.

## Key findings

- The proposed models achieved kappa values of 0.6 or higher, indicating substantial agreement in scan quality assessment.
- The ensemble system reduced the time lapse for image classification compared to the full sequence method.

## Abstract

Developmental Dysplasia of the Hip (DDH) is a relatively common hip joint disorders in infants, affecting one to three per a thousand births. If found early, it can be treated preemptively by simple non-invasive methods. But if not, then several surgical procedures may be required that can cause high economic burden. The accuracy of diagnosis using ultrasound (US) images heavily relies on locating anatomical landmarks on the image. However, there is an intra-observer/inter-observer variability in determining the exact location of the landmarks. In this study, an automated scan quality assessment system of pelvic US image by evaluating quality of five landmarks using transfer learning models was proposed.

US images from 1,891 subjects were obtained at two hospitals in the Republic of Korea (henceforth Korea). Also, an ensemble system was developed using transfer learning models to automatically evaluate the scan quality by scoring five anatomical landmarks. Gradient-weighted class activation mapping was used for verifying whether models that reflect the geographical features of the images had been properly trained. Considering the applicability in the real-time environment, this study proposes an alternative sequence method (ASM) that has been discovered to have improved the lapse of scan quality assessment.

All the selected models achieved kappa values of 0.6 or higher, indicating substantial agreement, and the AUC score for classifying standard images based on the total score was 0.89. The activation map of the trained models properly reflected the structural features of the image. The time lapse for standard image classification was 0.35 second per image in full sequence method, and that of the three versions - ASM-1, ASM-2, ASM-3 - were 0.27, 0.22, and 0.20, respectively.