Date of Award

2-28-2024

Publication Type

Thesis

Degree Name

M.A.Sc.

Department

Electrical and Computer Engineering

Keywords

Computational Time;Deep Learning;Ensemble Model;Machine Learning;ResNet101+VGG16;Skin Cancer

Supervisor

Jonathan Wu

Creative Commons License

Creative Commons Attribution 4.0 International License
This work is licensed under a Creative Commons Attribution 4.0 International License.

Abstract

Cancer is a leading global cause of mortality, with skin cancer posing a significant threat due to its high fatality rate. Early detection and intervention are paramount. Deep learning algorithms have shown promise in skin lesion classification, enhancing expert diagnosis. Our research introduces a novel approach, blending Machine Learning (ML) and Deep Learning (DL), featuring an ensemble model (VGG16 + ResNet101). This ensemble capitalizes on VGG16's local feature capture and ResNet101's advanced feature extraction. To optimize accuracy, our method employs pre-processing, including resizing, color correction, and edge enhancements. Segmentation, using the Otsu algorithm, isolates lesions, with noise removal for accuracy. The performance of our method is rigorously assessed across three distinct datasets: the International Skin Imaging Collaboration (ISIC) 2018, 2019, and 2020. This evaluation includes a comprehensive analysis of computation time and accuracy, comparing our method with other pre-trained deep learning models, such as ResNet50, VGG16, InceptionV3, and Xception. Our results demonstrate the superiority of our ensemble model, VGG16 + ResNet101, achieving impressive accuracy rates of 89.35\%, 90.43\%, and 93.26\% on the ISIC 2018, 2019, and 2020 datasets, respectively. In our ensemble model, we not only enhance accuracy compared to individual models but also manage to reduce computing time, addressing this critical trade-off that usually did not happen in recent research. The utilization of deep learning algorithms in our method reduces the influence of human-related variables during the diagnostic process, ensuring the delivery of reliable and consistent results. This research shows great promise in advancing skin cancer diagnosis and contributing to the field of medical image analysis.

Download

Share

COinS