Artificial Intelligence for Skin Lesion Analysis based on Computer Vision and Deep Learning

Alzahrani, Saeed
(2023) Artificial Intelligence for Skin Lesion Analysis based on Computer Vision and Deep Learning. PhD thesis, University of Liverpool.

[img] Text
201338546_Feb2023.pdf - Author Accepted Manuscript

Download (17MB) | Preview


Skin lesions appear in various sizes and forms and can be localised in one place or spread across the whole body due to different conditions. Dermatologists typically undertake physical examinations to diagnose skin lesions. However, this task costs time and requires excessive effort and can be inconsistent. Depending on the type of lesion and whether or not malignancy is present, additional diagnostic testing, such as imaging or biopsy, may be needed. Computer-aided diagnosis (CAD) systems, using clinical and dermoscopic images, could provide a quantitative assessment tool to help clinicians identify skin lesions and evaluate their severity. The recent progress in computer vision and deep learning has encouraged researchers to harness medical imaging data to develop powerful tools which could provide better diagnosis, treatment and prediction of skin conditions. By leveraging artificial intelligence techniques, including computer vision and deep learning, this work introduces intelligent computerised approaches using dermoscopic and clinical images to analyse and identify two types of skin lesions producing enhanced medical information. This thesis designed, realised, and evaluated the benefit of features learned automatically from images through the stacked layers of convolution filters in the convolutional neural network (CNN) models. The final objective of conducting the research in this thesis is to benefit patients with skin lesion condition assessment and skin cancer identification without adding to the already high medical costs. An automated regression-based method has been developed in this thesis for acne counting and severity grading from clinical facial images. In addition to the acne lesions, another type of skin lesion has been considered, represented by melanoma-related lesions. Two pipelines have been presented in this thesis to identify melanoma lesions. The first framework benchmarks and evaluates several CNN models for melanoma and non- melanoma classification from only dermoscopic images. While the second developed model for melanoma detection integrates the seven-point checklist scheme with CNN using both clinical and dermoscopic images. The experimental results of the work presented in this thesis manifest improved/ competitive performance compared to the state-of-the-art skin analysis methods using several evaluation metrics. The findings of the developed approaches demonstrated effective analysis of skin lesions with high accuracy, reducing the risk of misdiagnosis, and providing a more efficient means of detecting melanoma and automated acne lesion severity grading. Additionally, the application of computational intelligence allows for cost savings by reducing the need for manual analysis and enabling the automation of grading support, resulting in a more reliable and consistent process. Overall, the new automated methods based on computational intelligence demonstrate the benefits of developing computer vision and deep learning techniques for skin lesion analysis towards early skin cancer identification and cost-effective and robust grading support.

Item Type: Thesis (PhD)
Divisions: Faculty of Science and Engineering > School of Electrical Engineering, Electronics and Computer Science
Depositing User: Symplectic Admin
Date Deposited: 20 Jul 2023 14:07
Last Modified: 20 Jul 2023 14:07
DOI: 10.17638/03168503
  • Al-Nuaimy, Waleed
  • Al Ataby, Ali