Etudiant : DIBBA PA ALIEU

Filière : Machine Learning Avancé et Intelligence Multimédia (MLAIM)

Encadrant : Pr. SABRI ABDELOUAHED

Annèe : 2025

Résumé : This project tackles the persistent challenge of class imbalance in skin lesion datasets, with emphasis on improving melanoma detection. The HAM10000 dataset, which is significantly skewed toward benign lesions, serves as the primary data source. To address this imbalance, the study explores the use of Generative Adversarial Networks to generate synthetic melanoma images, thereby augmenting the dataset and ensuring more equitable representation of clinically critical classes. Two baseline Deep Convolutional GAN models were implemented and evaluated. The first model followed the standard architecture, utilizing BCELoss with Sigmoid activation, but suffered from instability and low quality image generation. To mitigate these issues, a second variant was introduced with several enhancements, including the use of BCEWithLogitsLoss, spectral normalization, the AdamW optimizer, and automatic mixed precision. These modifications led to improved training stability and a reduced Fréchet Inception Distance, although mode collapse still occasionally occurred with extended training. In the subsequent phase of the project, various attention mechanisms were integrated into the GAN architecture to further enhance the quality of generated images with respect to lesion boundaries, textures, and fine-grained visual features. These attention augmented GANs were then employed to produce synthetic samples for both binary (melanoma vs. non-melanoma) and multi-class classification tasks. For the multi-class setup, a conditional DCGAN (cDCGAN) was used, enabling class specific image generation across all six underrepresented diagnostic categories in the HAM10000 dataset. The resulting experiments evaluated the impact of these synthetic images on classification performance, demonstrating the potential of attention enhanced generative models to mitigate data imbalance and improve diagnostic accuracy in both binary and multi-class skin lesion classification settings.