A Study of Convolutional Neural Network Architecture for Hip Fracture Diagnosis Using Radiographs on Small Datasets

Currently, there are insufficient doctors in Thailand. According to a report on the population of Thailand, the ratio of doctors to the population is 1: 1,135, which is lower than the World Health Organization's (WHO) recommended ratio of 71: 1,000. It indicates that Thailand has a shortage of healthcare personnel. As a result, the population is unable to receive effective treatment. Hip fractures are an ordinary problem among aging individuals. Outpatients are typically treated in an emergency room without a professional specialist. The issue of inaccurate diagnosis and deferred treatment worsens the elevated risk of complications and casualty rate. The issue of wrong conclusions and deferred treatment worsens the lifted hazard of complications and casualty rate. The use of Convolutional Neural Networks (CNN) can assist physicians with limited experience in diagnosing anterior-posterior (AP) pelvic orthotic X-ray images of hip fractures quickly and confidently. This paper reports a study of applying image preprocessing to enhance the diagnosing performance of CNN on small datasets. VGG16 and DenseNet-121 were used in the study. The results showed that the proposed image preprocessing framework could increase the diagnostic accuracy of VGG and DenseNet-121 by 24.23% and 24.94%, respectively.