Browsing by Author "Friedling, Louise"

Now showing 1 - 2 of 2
  • Thumbnail Image
    Item
    Open Access
    Ancestry-independent osteometric sex estimation from selected postcranial skeletal elements of South Africans: a machine learning approach
    (2022) Scott, Michelle; Friedling, Louise
    Sex estimation, as part of a biological profile, has the power to halve the number of possible identities of unidentified skeletal remains. Postcranial elements have been studied in South Africa (SA) for the purpose of sex estimation and have often proven to be more accurate than the cranium. Estimation techniques using postcranial elements in SA almost exclusively utilise discriminant analysis to evaluate sex, but international publications have shown success using alternative machine learning (ML) algorithms. SA methods and standards are often restricted by limited sample size, lack of robust statistical techniques in older publications and, the prerequisite of known or estimated ancestry. Most methods are specific to SA African, European or, more recently, Mixed ancestry groups and are unreliable when ancestry is unknown. The aim of this study was to apply a series of ML algorithms to train ancestry independent sex classification models using postcranial osteometric measurements from the cadaveric skeletal remains of modern South Africans, focussing on long bone joints. The study consisted of a roughly demographically representative, pooled sample, of 650 South Africans (325 male, 325 female). 12 osteometric measurements were taken from available left- and, or right-sided bones for each individual. All 12 mensurations were sexually dimorphic and differences between left- and right-sided bones were negligible. The dataset was subjected to ML algorithm training using univariate and multivariate predictor combinations. The best performing ML algorithm, given the sample size and available predictors was discriminant function analysis. Univariate model accuracies ranged from 80.5-89.1% and multivariate model accuracies ranged from 84.5%, using 2 predictors, to 92.8%, using 12 predictors. An optimised 3-predictor model was able to predict sex with 92.7% accuracy. Results from this study were comparable to those using ancestry-specific models and non-ancestry-specific models, where available. Findings from this study suggest that the inclusion of ancestry, when predicting sex using the elements examined, is not necessary as it does not significantly improve prediction accuracy.
  • No Thumbnail Available
    Item
    Open Access
    Bone age estimation using machine learning approach: An assessment using hand and wrist bones of South African children
    (2023) Nam, Yuseung; Friedling, Louise
    Skeletal maturation is influenced by various factors such as genetics, hormonal secretions, and nutrition. Establishing a skeletal maturity level in children becomes necessary when a deviation from the standard growth patterns may indicate signs of diseases; and whether that individual is a minor. Bone Age Assessment (BAA) achieves this, as it is a clinical process used to establish an individual's biological profile. A large proportion of the South African population resides in rural areas where the fully functional civil registration system is limited. Many individuals remain unregistered on the national database, bringing about various challenges. This reduces the likelihood of unregistered children receiving favourable treatment in judicial cases or access to amenities at juvenile rehabilitation centres. Moreover, it puts them under the same threat of abuse and discrimination as adult offenders. Typical clinical methods for BAA are the Greulich and Pyle (GP) and Tanner and Whitehouse (TW) methods using wrist radiographs of the left-hand. Although these methods have been updated throughout the decades, they rely on experienced radiologists' manual power, which is highly time-consuming, resulting in intra- and inter-observer errors. Our study uses a machine learning method to train and automatically predict bone age with carpal bones from a sample of South African children to mitigate these problems. Two datasets of 12,611 North American population (RSNA) and 400 South African population (SA) left-hand X-ray radiographs (from a LODOX machine) were used from birth to 19 years of age. These radiographs of the two datasets were pre-processed to remove unnecessary labels, remove the background, and straighten the X-ray image. The first experiment used the pretrained models, Xception, InceptionV3, MobileNet, and VGG-16, using the pre-processed and unprocessed datasets and comparing their performance. The pre-processed dataset was selected for model benchmarking to find the best-performing model for bone age estimation out of the four pre-trained models. Scatterplots of the four models were plotted to visualise their generalisation performance on bone age estimation. Xception was the best-performing bone age model used to determine bone age prediction using combined RSNA and SA datasets as train sets. Due to the overwhelming difference in sample sizes between RSNA and SA datasets, imbalanced and balanced data training was applied to overcome the difference. The best-performing model - Xception, achieved a mean absolute error of 5.70 months when using population-specific pre-processed data. Bone age estimation benefits more from a machine learning model than a simple linear regression model when using a raw X-ray image input. The combined RSNA (10,000) + SA (300) train set of the Xception model achieved an MAE of 7.43 months from RSNA and 14.36 months from the SA dataset. The results suggest that bone age estimation using different populations as train and test sets contributes to less accurate bone age prediction, indicating a need for a population-specific model. The imbalanced and balanced data training proved that more samples for the South African population are needed for accurate bone age prediction, as bone age MAE decreased with an increasing number of minority SA datasets increased. The population-specific bone age model significantly outperformed the manual methods. The population of South Africa is diverse and distinctive, with a wide range of ancestral and genetic backgrounds that might impact bone growth. Future studies should focus on creating a bone age estimation model tailored to this unique population.