Date of Award


Publication Type


Degree Name



Electrical and Computer Engineering

First Advisor


Second Advisor


Third Advisor

P.Moradian Zadeh


Computer-aided diagnosis, Deep learning, Generative adversarial network, Melanoma detection, Skin lesion classification, Soft attention



Creative Commons License

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


The rarity of Melanoma skin cancer accounts for the dataset collected to be limited and highly skewed, as benign moles can easily mimic the impression of the melanoma-affected area. Such an imbalanced dataset makes training any deep learning classifier network harder by affecting the training stability. We have an intuition that synthesizing such skin lesion medical images could help solve the issue of overfitting in training networks and assist in enforcing the anonymization of actual patients. Despite multiple previous attempts, none of the models were practical for the fast-paced clinical environment. In this thesis, we propose a novel pipeline named SkinCAN AI, inspired by StyleGAN but designed explicitly considering the limitations of the skin lesion dataset and emphasizing the requirement of a faster optimized diagnostic tool that can be easily inferred and integrated into the clinical environment. Our SkinCAN AI model is equipped with its module of adaptive discriminator augmentation that enables limited target data distribution to be learned and artificial data points to be sampled, which further assist the classifier network in learning semantic features. We elucidate the novelty of our SkinCAN AI pipeline by integrating the soft attention module in the classifier network. This module yields an attention mask analyzed by DenseNet201 to focus on learning relevant semantic features from skin lesion images without using any heavy computational burden of artifact removal software. The SkinGAN model achieves an FID score of 0.622 while allowing its synthetic samples to train the DenseNet201 model with an accuracy of 0.9494, AUC of 0.938, specificity of 0.969, and sensitivity of 0.695. We provide evidence in our thesis that our proposed pipelines outperform other state-of-the-art existing networks developed for this task of early diagnosis.