Development of a new MRI hand/wrist living age estimation methodology
Research type
Research Study
Full title
Development of a new methodology for living age estimation based on the analysis of magnetic resonance images of hand/wrist of a Scottish population
IRAS ID
262576
Contact name
Lucina Hackman
Contact email
Sponsor organisation
University of Dundee
Clinicaltrials.gov Identifier
Sponsor Ref number, 2-059-19
Duration of Study in the UK
2 years, 4 months, 31 days
Research summary
The present research will analyse magnetic resonance images of the hand/wrist area, focusing on skeletal development, with the final aim of developing a new methodology for living age estimation. Living age estimation is the forensic practice where the age of an individual, who lacks valid identity documents, is estimated for juridical issues, e.g. cases related to unaccompanied minors applying for asylum in European countries.
Currently, the main medical imaging method upon which age estimation relies is radiographic analysis. Nevertheless, recently, the use of x-rays for living age estimation has become an ethical and legal issue of international relevance. As stated by the European Asylum Support Office (Second Edition, 2018), some countries, e.g. France, have restricted the use of ionising radiation examination for age estimation for legal purposes. Therefore, the need to develop new methodologies based on magnetic resonance imaging (MRI) which does not involve ionising radiation. Furthermore, it is known from previous studies on MRI, that this technique can also depict in detail anatomical structures which are relevant to age estimation, such as the skeletal development and growth plate fusion of the hand/wrist area.
Regarding sample size and collection, both female and male subjects will be included, with an age range of 10-26 years. Since this research is a non-clinical trial, the magnetic resonance images will be collected from the pre-existing database of Tayside NHS Trust at Ninewells Hospital.
The study will last for four years: the first part of the research will be focused on the literature review, the second part on the collection and the analysis of magnetic resonance images and the third part to the development and testing of a new methodology based on the information gathered from the analysis of the MRI.Summary of results
The title of this study is ‘Development of a new methodology for living age estimation based on the analysis of magnetic resonance images of hand/wrist of a Scottish population’. The study was carried out by the Principal Investigator, Dr Valentina Panci, as part of her doctoral research, with the supervision of the Chief Investigator, Professor Lucina Hackman, and Dr Catriona Davies. The Sponsor of the study was Dr Vera Nuritova at the TASC research & Developmental Office.The study aimed to develop a method to estimate a person's age for forensic purposes (living age estimation). This is useful to determine the age of children or people when their exact birthdate is unknown. There are different reasons for which an age estimation is needed, such as legal and judicial purposes, unaccompanied minors, adoptions, and sports and competitions.
Over the years, researchers have created different methods to estimate a person's age by observing how different parts of their bones develop and how this relates to their chronological, age. Traditionally, radiographies have been the go-to medical method for these age estimations. However, radiographies use ionising radiation, which has raised concerns about whether it is ethical to expose children to this kind of radiation when it is not necessary for their health. Because of these concerns, researchers are now focusing on using magnetic resonance imaging (MRI), which does not use any radiation and is safer for children. In this study, as part of the method, it was also decided to use traditional machine learning algorithms, specifically Decision Tree and Random Forest with regression analysis, to estimate the age of the participants through the observation of the bones’ growth of the wrist, specifically of the radius and the ulna bones.
The research objectives of this study were to:
• Identify specific bone features that are most closely related to age.
• Determine which machine learning algorithm, between Decision Tree Regressor or Random Forest Regressor, is more accurate for age estimation.
• Check if including the segmentation of certain bone areas (distal diaphyses) improves age estimation accuracy.
• Compare whether the algorithms perform better using single slices of images (experiments by slice) or using the combined areas from all slices for each individual (experiments by individual).
• See if there is a difference in age estimation accuracy between boys and girls, and find out which algorithm gives the best precision for each sex.This study used only pre-existing MRI scans from the Tayside NHS Trust’s database, so no direct contact with the patients was needed at any stage. The data was collected in May 2021. The magnetic resonance images (MRIs) were accessed and collected from the Picture Archiving and Communication System (PACS), using computers available at the Department of Radiology at Ninewells Hospital in Dundee. Then, the Principal Investigator at the University of Dundee, analysed the magnetic resonance images and developed a new method for age estimation.
The study included two types of MRI scans that are called T1-weighted and Proton density showing the bones of the hand and wrist, in particular the radius and the ulna bones. These scans were collected from both boys and girls with an age between 10 years and 26 years. For the T1-weighted MRI scans, a total of 125 were collected from girls, and a total of 79 from boys. For the Proton density MRI scans, a total of 178 were collected from girls, and a total of 99 were collected from boys.
The methodology involved different steps. The first step included the standardisation of the images. To make the images consistent, only MRIs showing both radius and ulna bones were kept. Then the images were all cropped at the same size using Adobe Photoshop, and saved with a format called TIFF, which keeps the quality high without losing data. Next, the Image Labeler app of a program called MATLAB, was used to create labels and highlight (segment) the pixels areas of the bones necessary for the analysis. Four labels were created: one for the epiphyseal area of the radius bone, one for the epiphyseal area of the ulna bone, one for the epiphyseal gap area of the radius bone, and finally one for the epiphyseal gap area of the ulna bone.
Then, a MATLAB script was created to generate masks from the segmentation of the pixels of the images. This allowed to calculate and extract the areas of the growth plates and the ends of the radius and ulna bones (called epiphyseal gaps and epiphyses). The ends of the radius and ulna were segmented and measured only for individuals whose growth plates were not fully fused. This was done to investigate if including these areas could improve the accuracy of the age estimation methods. This process was completed for each one of the datasets, the T1-weighted MRI scans from girls, the T1-weighted MRI scans from boys, the Proton density MRI scans from girls and the Proton density MRI scans from boys.
Finally, two machine learning algorithms, Decision Tree and Random Forest with regression analysis, were created with the aim to estimate the age using the data collected from the segmentation of the images (areas of the growth plates and the epiphyses). For each dataset, the areas were organised in four different excel files. In the first excel file, the areas of the growth plates were organised by MRI slice. In the second excel file, the areas of the growth plates were organised by individual. In the third excel file, the areas of the growth plates and the areas of the epiphyses were organised by slice. Finally, in the fourth excel file, the areas of the growth plates and the areas of the epiphyses were organised by individuals. This organisation helped in using the data efficiently to train the models and investigate if the algorithms perform better using single slices of images (experiments by slice) or using the combined areas from all slices for each individual (experiments by individual). Thirty-two Experiments in total were carried out, 16 with the Decision Tree algorithm and 16 with the Random Forest algorithm. The data of each dataset were divided in a training set, to train the algorithms, and a test set, to test the algorithms on unseen (new) data.
For the analysis, it was investigated when (at what ages) the radius and ulna bones started to fuse (timing of active fusion). For the T1-weighted MRI scans of girls, the T1-weighted MRI scans of boys, and the Proton density MRI scans of boys, the ages when radius and ulna started to fuse were in agreement with previous studies. For the Proton density MRI scans of girls, in two cases beginning of fusion was first recorded for a girl aged 13 years for the radius, and a girl aged 12 years for the ulna. These were two isolated cases, and the active fusion was observed only in one slice of the MRI scans. The first girl to show active fusion in more than one slice was 14 years old, age that was in agreement with those recorded in previous studies.
For the complete fusion of the radius for the T1-weighted MRI scans of boys, the Proton density MRI scans of boys, and the Proton density MRI scans of girls, the results were in agreement with the literature. For the T1-weighted MRI scans of girls the complete fusion of the radius was found at 18 years of age, which was in disagreement with the previous studies using MRIs. For the complete fusion of the ulna for the T1-weighted MRI scans of boys and T1-weighted MRI scans of girls the age of complete fusion was in agreement with previous studies. For the Proton density MRI scans (both of girls and boys), no comparison was possible since the lack of information of the fusion of distal ulna using this type of MRI scan in the literature.
Two analyses were also done to investigate which features (areas of the growth plates or areas of the epiphyses) were most important for estimating the age (Feature Importance Analysis and Pearson’s Correlation Analysis). In the Experiments where each MRI slice was analysed, the area of the epiphyseal gap of the radius bone was the most important factor and was mostly correlated with age. In the Experiments where all data were analysed by individual, the results were more difficult to interpret, and some inconsistencies were noticed. These might be due to the fact that the number of MRI slices was not the same for each person. Since the MRIs used in this study were collected from a pre-existing database, it was not possible to standardise the number of slices for all MRI scans. Caution should be applied when interpreting these findings.
About the two machine learning algorithms, both gave more accurate results when the areas given to the algorithm were organised by individual, suggesting that during the training phase of the algorithms, all the slices for each person should be taken in consideration. This allows the machine to train through the observation of all the features of each bone, leading to more accurate results.
The Decision Tree algorithm did not perform well on the test set, especially for the T1-weighted MRI scans from boys. This poor performance is due to a phenomenon in machine learning called "overfitting." Overfitting means that the algorithm learns to perform very well on the training set (the data it was trained on), but it fails to work as well on the test set (new, unseen data). The Random Forest algorithm was overall more accurate on the test set, with the best results obtained when all the features were taken in consideration (areas of the epiphyseal gaps and of the epiphyses of both radius and ulna). This suggests that using this combination of features would be the best approach for estimating a person's age. Additionally, the Random Forest algorithm was more accurate at estimating the age using data from the T1-weighted MRI scans of both girls and boys. This suggest that this type of MRI scan should be the preferred for age estimation. However, many factors could influence these results. Currently, there is not enough information in the literature about using Proton density MRI scans of the hand and wrist for age estimation, indicating that further research is needed.
The small numbers of MRI scans in the datasets might have influenced the results.
The non-standardisation of the number of slices for each MRI scans, might also have played a role, especially in the evaluation of the results of the most important feature for the estimation. Additionally, the areas of the carpal bones of the hand were not included in this study. However, the inclusion of the areas of these bones in the analysis could help with the improvement of the accuracy of the algorithms especially in the younger ages.For future research, more studies should be conducted to explore how machine learning can be used for estimating age. This includes using larger datasets with more individuals from each age group to test and refine the methods. The differences observed when estimating age based on individuals versus MRI slices should also be further investigated. This will also help to better understand which features are most important and how they relate to age, making machine learning methods clearer and more reliable. Additionally, including the carpal bones in the analysis could be useful. This could help determine if these features improve age estimation accuracy. Finally, experimenting with classification tasks, rather than regression, could be beneficial, to investigate the differences between classification and regression systems for age estimation, and to help train the algorithms to assign individuals to specific age groups.
REC name
East Midlands - Derby Research Ethics Committee
REC reference
20/EM/0119
Date of REC Opinion
14 May 2020
REC opinion
Favourable Opinion