Open Collections

Du, Siyi

University of British Columbia

Skin cancer is the most common cancer worldwide, accounting for a third of all cancers. A major public health problem with shortages of dermatologists and high rates of misdiagnosis, Computer-aided diagnosis (CAD) offers much for easing the workload on clinicians and reducing human error. Deep neural networks (DNNs) have already shown promise in this domain and with impressive results in skin lesion segmentation (SLS) and skin lesion classification (SLC), both crucial steps in dermatology CAD systems. However, current approaches are still plagued with critical limitations that impede their full potential in real-life clinical applications. A main problem is their high susceptibility to data bias, which manifests as problematic unfairness in their decision-making. In dermatology, this is manifested with variable levels of model accuracy across skin tones (types). Another main problem relates to more recent DNNs that are based on vision transformer (VIT), in which, diverse, yet small-sized skin databases. Fare poorly due to the inherent nature of data-hungry models. In this thesis, we made contributions to both SLS and SLC. First, we proposed FairDisCo, a disentanglement DNN with contrastive learning, which adds a dedicated ‘fairness’ network branch that reduces sensitive attributes (skin-type information) from model representations, namely, and another contrastive branch to improve representation learning for better diagnosis accuracy. Second, we proposed MDViT, a multi-domain VIT with domain adapters to mitigate model data-hunger and to combat negative knowledge transfer (NKT) that decreases model performance on domains with inter-domain heterogeneity. MDViT also employs mutual knowledge distillation to enhance representation learning across domains. Third, we proposed AViT, an efficient framework that utilizes lightweight modules within the transformer layers to transfer a pre-trained VIT to the SLS task without a need for updating its pre-trained weights for data-hungry mitigation. AViT also employs a shallow convolutional neural network (CNN) to produce a prompt embedding with fine-grained information and to inject the CNNs’ inductive biases for better representation learning. All proposed models were thoroughly tested on publicly available databases and validated against state-of-the-art (SOTA) algorithms with comprehensive quantitative results demonstrating superior performance, both in terms of accuracy, fairness, and efficiency.

Text

2023-09-14

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/85920

Electrical and Computer Engineering

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/