Dermoscopy-based skin lesion analysis is an important computer-aided diagnosis task because early identification of malignant lesions can support timely clinical intervention. However, deep learning models trained on dermoscopic images are often affec...
Dermoscopy-based skin lesion analysis is an important computer-aided diagnosis task because early identification of malignant lesions can support timely clinical intervention. However, deep learning models trained on dermoscopic images are often affected by class imbalance, particularly when benign melanocytic nevi dominate the training distribution while clinically important minority categories are underrepresented. This thesis addresses this problem using generative data augmentation for the HAM10000 skin lesion dataset. The study proposes Guided Fusion Auxiliary Classifier GAN (GF-ACGAN), an inpainting-guided class-conditional generative framework for dermoscopic image synthesis. GF-ACGAN combines a binary lesion mask and an inpainted dermoscopic image through channel-wise guided fusion, while an auxiliary classifier discriminator encourages class-consistent image generation. The proposed generator is compared with five baseline GAN configurations: Vanilla GAN, Conditional Wasserstein GAN with Gradient Penalty and FFT loss, FastGAN, LightweightGAN, and Auxiliary Classifier GAN. The resulting synthetic images are then used in downstream skin lesion classification experiments.
Experiments are conducted on HAM10000 across seven lesion categories: melanocytic nevus, melanoma, benign keratosis-like lesion, basal cell carcinoma, actinic keratosis/intraepithelial carcinoma, vascular lesion, and dermatofibroma. Image synthesis quality is evaluated using Fréchet Inception Distance, Kernel Inception Distance, Inception Score, and Structural Similarity Index. The experimental results identify GF-ACGAN as the strongest synthesis model in terms of FID, KID, and SSIM, achieving FID = 70.08, KID = 0.0351, and SSIM = 0.684. Downstream classification is evaluated using EfficientNetV2-M, SwinV2-Base, ResNet50, and ConvNeXtV2-Base under both image-wise and lesion-wise splitting protocols, with performance measured on a held-out fixed test set using test accuracy, balanced accuracy, macro precision, and macro F1. The classification results show that hybrid augmentation using traditional transformations and GF-ACGAN synthetic images is more effective than using real-only data or GAN-only augmentation in most settings. Under the image-wise split, SwinV2-Base achieves the best macro F1 score of 0.894 using a 75:25 traditional augmentation to GAN ratio. Under the stricter lesion-wise split, SwinV2-Base achieves the best macro F1 score of 0.832 using a 50:50 ratio. EfficientNetV2-M, ResNet50, and ConvNeXtV2-Base also show improvement over their Real- only baselines under the selected hybrid settings. These findings indicate that guided synthetic augmentation can improve model generalization under the evaluated settings for imbalanced dermoscopic classification when synthetic images are combined with traditional augmentation rather than used as a complete replacement.
Keywords: skin lesion classification, dermoscopy, HAM10000, class imbalance, GAN, ACGAN, GF-ACGAN, synthetic augmentation, inpainting, deep learning.