Use this URL to cite or link to this record in EThOS:
Title: Innovative lesion modelling for computer-assisted diagnosis of melanoma
Author: Liu, Zhao
Awarding Body: University of the West of England, Bristol
Current Institution: University of the West of England, Bristol
Date of Award: 2012
Availability of Full Text:
Access from EThOS:
Malignant melanoma is a relatively rare disease, but it is the most fatal form of skin cancers. In UK alone, it represents 10% of all skin cancers and its incidence has increased four times during the last three decades. The survival rate of melanoma is inversely proportional to the tumour thickness, so early detection with surgical removal of thin lesions is vital for successful treatment. Many dermatologists have advocated the development of computer- assisted diagnosis (CAD) for early detection of melanoma, due to its objectivity and the potential it could provide for self-screening. Numerous computer-based methods have been developed for melanoma diagnosis. However, better accuracy and robustness are demanded for computer-based systems to be trustworthy enough for routine clinical applications. With the aim of improving the existing CAD systems, the present study develops several innovative techniques to achieve accurate and reliable computer-based diagnosis for melanoma. Based on the clinical evidence, a new sub-segmentation scheme for cutaneous lesions is firstly proposed to allow the isolation of normal skin, suspicious lesion areas, and interesting darker areas inside the lesion, simultaneously. This scheme is much different from traditional segmentation techniques, where only lesion areas and non-lesion areas are separated. It has been found that the isolated darker areas within the lesion are of great diagnostic importance, such that they are useful in characterising the malignancy of cutaneous lesions. Melanin index and erythema index, which respectively characterise the amount and distributions of melanin and haemoglobin components within human skin, are computed from conventional RGB images according to the optical theory of human skin. These biological indices are employed to generate new colour variegation descriptors for melanoma diagnosis. Experiments show that the derived chromophore indices can accurately describe pigmentation distributions, and tend to be less influenced by external imaging complexities (e.g. light conditions and optical sensor parameters) than the conventional image intensities such as RGB colour values. A novel asymmetry analysis algorithm based on the global point signatures (GPSs) is developed to quantify the shape and pigmentation asymmetry of cutaneous lesions, simultaneously. In contrast to existing methods for asymmetry analysis, the newly proposed method results in only one pair of reflective symmetry axes. This is consistent with the asymmetry of cutaneous lesions as perceived by the human eye, and circumvents the problem of yielding two different pairs of symmetry axes when shape and colour asymmetries are evaluated separately. In addition, the new asymmetry descriptor is approved to be invariant to rigid transformations, and robust to non-rigid deformations. This suggests that the GPS-based asymmetry descriptor not only benefits for the computer-assisted diagnosis of melanoma, but also has good potential for follow-up monitoring of suspicious cutaneous lesions for high-risk Caucasian population. Finally, an innovative CAD system for melanoma is developed using the extended Laplacian Eigenmap. This system incorporates, for the first time, clinically important metadata into a completely automatic classification process for melanoma diagnosis. Algorithm performance is evaluated on both 2D dermoscopy images and 3D data obtained from our Skin Analyser device. The proposed CAD system achieves 90.97% sensitivity and 86.42% specificity with the dermoscopy database, and 94.12% sensitivity and 88.99% specificity with the Skin Analyser database. The diagnostic accuracy obtained by our system is superior to most of the results from other existing CAD systems for melanoma.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available