A practical application of scientific and engineering methodology to a real-world problem
Andrey Gizdov is the SIM Young Scientific Instrument Maker: 2019 and ISEF entrant. Here is his project entry -
The human skull is made by multiple plates (bones), connected together to form a solid structure. The connections between those plates are called cranial sutures and are located at the edges of the separate bones which form the skull. At an early stage of life, to allow for brain growth, the space between the edges of the neighboring plates (the sutures) are wide and filled with a flexible soft tissue. As the individual grows up to adolescence, the plates of the skull fuse together, forming a much narrower suture path and providing the skull bones with a more solid connection over time. This change in width and shape of the sutures is directly dependent on aging, as it continues throughout the life of the developing human body.
Forensic scientists and anthropologists have been using this dependency for many years: to assess the degree of fusion along the sutures and estimate the Age-At-Death (AAD) of the skull’s holder. Until now, assessing the suture has relied on a ranking system based only on the opinion of the particular researcher, without taking into consideration any objectively measured metrics. Besides introducing a lot of subjectivity and inaccuracy in the assessment, this is also incredibly time consuming to do on bulk quantities of skull data. This project addresses the problem, by elaborating an algorithm for more accurate and objective AAD estimation, based on assessment of the suture cross- section with the help of mathematics and machine learning.
For the purpose of obtaining suture data, I’ve used volumetric images of dry skulls, generated by an industrial μCT system. The first part of the project deals with the autonomous extraction of cross-sectional suture images at perpendicular to the skulls surface incline, which used to be done by hand until now. The obtained spatial resolution (voxel size of 97.5 μm for μCT) of the extracted images was high enough to allow precise detection of the suture using a semantic segmentation neural network with an accuracy of 92%.
The pixel color intensities in the segmented suture area are split between ‘light’ and ‘dark’ using adaptive thresholding and C-Means clustering. An unfused suture will contain a lot of empty space between the neighboring bones, which is represented by black color on the μCT scans. Respectively, a fused suture will have a large part of its area solidified into bone matter, which is represented by lighter colors. Calculating the ratio between the number of ‘dark’ and total number of pixels, provides useful information about the degree of closure. An unfused suture will have most of its pixels classified as ‘dark’ due to the large empty space, resulting in this ratio being close to 1. Alternatively, a fused suture has most of its pixels turned into white bone matter, which results this value being closer to 0.
The ratio in the example mentioned above, and several other metrics, have been recorded by the algorithm for a population of male skulls with known AADs. Using statistical software and outlier filtering, I’ve found a regression equation which describes the Age-At-Death variable in terms of the measured by the algorithm metrics. I’ve tested the prediction accuracy of the regression on an independent sample of skulls and managed to reduce the error of the predicted AADs over 3 times compared to the already existing methods for AAD estimation based on cranial suture analysis.
The project saves human resources by completely automatizing the process of cranial suture assessment and provides a significantly more accurate Age-At-Death prediction than any of the existing methods in the field. Overall, this major improvement is an important advancement in the fields of forensic science and archaeology, as it introduces a new objective and automatic method for Age-At-Death estimation in cases of unidentified skeletal remains.