Document Type

Theses, Ph.D

Disciplines

2. ENGINEERING AND TECHNOLOGY, 2.2 ELECTRICAL, ELECTRONIC, INFORMATION ENGINEERING

Abstract

Image-based AI systems that analyse human skin are increasingly used in healthcare and computer vision applications. However, many human skin-based image datasets do not provide reliable information about skin type, making it difficult to assess whether these systems perform consistently across the full spectrum of skin colour. The objective of this thesis is to examine how skin type diversity is represented and measured in image datasets, and to evaluate the reliability of image-based skin type measurement methods under different imaging conditions. Using publicly available skin lesion image datasets as a well-defined and widely used sub-class of skin image datasets, this research first conducted a systematic investigation to assess the availability, diversity, and distribution of skin type metadata. It then reviewed existing skin type measurement and classification approaches, highlighting their practical and conceptual limitations when applied to datasets. The performance of the widely used Individual Typology Angle (ITA) method was evaluated using dermatologist labelled datasets captured under varying lighting conditions. In addition, a new skin image dataset was designed and collected under controlled lighting conditions, incorporating both image data and objective colourimetric measurements obtained using a colourimeter. The results showed that the reliability of ITA-based skin type measurement is dependent on imaging conditions. Under controlled and consistent lighting, ITA demonstrated good agreement with physical colourimetric measurements, whereas variations in lighting conditions lead to increased variability and inconsistent skin type estimates. In contrast, physical colourimetric measurements remained stable across lighting conditions, highlighting condition-dependent discrepancies between image-based and ground-truth values. The key contributions of this thesis include: (i) a systematic investigation of publicly available skin lesion image datasets to examine the availability, quality, and diversity of skin type metadata; (ii) a review of existing skin type measurement and classification methods, highlighting their practical and conceptual limitations when applied to image datasets; (iii) an empirical evaluation of performance of the ITA method using existing dermatologist-labelled datasets under varying imaging conditions; and (iv) the design and creation of a new skin image dataset captured under controlled lighting conditions, providing skin type information through objective colourimetric measurements

DOI

https://doi.org/10.21427/66ny-sv27

Creative Commons License

Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License
This work is licensed under a Creative Commons Attribution-NonCommercial-Share Alike 4.0 International License.

Download


Share

COinS