Robust Statistical and Artificially Intelligent Approaches for the Analysis of 2D and 3D Morphological Data

  1. Courtenay, Lloyd Austin David
Supervised by:
  1. Diego González Aguilera Director
  2. José Yravedra Sainz de los Terreros Co-director

Defence university: Universidad de Salamanca

Fecha de defensa: 09 February 2023

  1. Julia Aramendi Chair
  2. Susana del Pozo Aguilera Secretary
  3. Francesco Boschin Committee member

Type: Thesis


Notions of geometry and morphology are a fundamental component of the way we perceive, describe, and essentially interact with objects. The shape and size of an element can be highly informative, and will thus condition the way we carry out basic functions in our daily lives. Morphology can be useful for; the detection of anomalies and patterns, the characterisation of an object or organism, as well as the identification of casuality (cause and effect). Nevertheless, finding an efficient and objective means of characterising morphology in both micro and macroscopic elements is often a great challenge in many fields of science. While many approaches to these types of tasks have long relied on a visual or qualitative description of shape and form, unfortunately most of these methods are influenced by notable degrees of subjectivity, typically product of human-based error and dependent on experience and perspective. In the present Doctoral Thesis, a wide array of different techniques for the extraction and analysis of morphological data is explored and discussed. Specifically, the main goals of this study are to define a general workflow that can be used for the quantification of different elements, with the hope of developing a transparent and transdisciplinary approach that can be applied to many fields of science. For this purpose, multiple techniques for the digitisation of both 2D and 3D data have been used, including (in order of prevalence); structured light surface scanning, micro-computed tomography, 3D-digital microscopy, and traditional photography for clinical imaging. Investigation into the best means of extracting morphological information was then explored, primarily including geometric morphometric analyses, while also testing combinations of traditional metric and elliptic Fourier analyses as well. Following this, the present Doctoral Thesis dedicates a significant portion of research into finding the best means of statistically analysing this type of information. Here the use of multiple robust statistical approaches is employed, as well as both parametric and non-parametric testing, in order to obtain the highest possible accuracy in the conclusions withdrawn. So as to leverage this information for tasks such as classification or diagnostics, this Doctoral Thesis also focuses great attention on the use of Computational Learning for the development of Artificially Intelligent algorithms in decision-making tasks. Prior to the development of classification algorithms, however, research delved into the possible limitations present in data science. Namely, problems due to sample size and the "curse of dimensionality". So as to overcome these limitations, different research lines were developed; first exploring the multitude of available techniques for data augmentation and simulation, followed by experimentation with different types of algorithms for classification tasks. The results obtained from this study reveal many different techniques to be useful for the modelling, extraction, and study of morphological information. Here it has been shown in a variety of different scenarios how, especially when combined with robust statistical approaches, both geometric morphometrics and elliptic Fourier analyses are powerful tools for the description of shape and form. Throughout this research, data simulation has also proven to be a fundamental step in the workflow, providing Artificially Intelligent Algorithms such as Neural Networks and Support Vector Machines sufficient information for the identification of new specimens. So as to promote transparency and improve reproducibility, the present doctoral thesis is also accompanied by a large collection of open-source code, datasets, and different software. The applicability of these methodological approaches, and thus their transdisciplinary nature, has been demonstrated across multiple case studies. These include applications in archaeological and palaeontological taphonomy, palaeoanthropology, animal wellfare, and dermatology. Through this compendium of research articles, the presented methods have been able to discover a number of different features from each of these fields. These range from the ability to identify extinct carnivore taxa in archaeological sites based on their tooth marks, to the first empirical quantification of skin lesion asymmetry as a diagnostic tool in dermatological oncology. In addition, through the presentation of a new mathematical model for the description of morphology, this study has been able to provide a new, more efficient, means of extracting biomechanical information from great primate limb long bones. Each of these discoveries present promising advantages for the study of other types of morphological data as well. The present Doctoral Thesis thus hopes to provide a new perspective on the means in which the morphology of different elements can be studied, promoting a more robust, transdisciplinary approach. Through this, future research will focus on applying these techniques to other fields of science, while working on fine tuning this methodological workflow to obtain higher precision and accuracy.