Advancement In Skin Lesion Analysis Using Deep Learning Methods

Skin cancer represents a major global health challenge, with steadily increasing incidence rates and substantial economic and clinical burdens associated with diagnosis and treatment. Among the various forms of skin cancer, melanoma is the most aggressive and life-threatening, accounting for a disproportionate number of skin cancer related deaths. Early detection is critical, as melanoma is highly treatable in its early stages but rapidly becomes fatal once it progresses. These challenges motivate the development of efficient, accurate, and scalable diagnostic tools capable of supporting clinical decision making. In this context, deep learning has emerged as a promising approach for automated analysis of dermatological images. This thesis investigates the application of deep learning techniques for the classification, generation, and segmentation of skin lesion images, with a particular focus on melanoma. The overarching objective is to improve model performance and reliability by examining key factors in dataset design and preprocessing, including data curation, synthetic data generation, and the influence of lesion background on classification. By addressing these factors, this work aims to enhance both the robustness and generalizability of deep learning models for skin cancer analysis. The first part of the thesis examines the effect of duplicate and nearduplicate images on skin lesion classification using convolutional neural networks (CNNs). Publicly available datasets, such as those from the ISIC archive, are widely used in dermatological image analysis but often contain redundant images that can bias training and inflate reported performance. Multiple CNN models are trained using datasets curated with different duplicate removal threshold indices to systematically evaluate the impact of redundancy. The results demonstrate that removing duplicate images leads to more reliable performance estimates and improved generalization, highlighting the importance of careful dataset curation in medical image analysis. The second part of the thesis focuses on synthetic dermatological image generation using generative adversarial networks (GANs). Data scarcity, particularly for rare skin conditions and underrepresented melanoma subtypes, remains a major limitation in training robust deep learning models. To address this challenge, this research proposes a Combined Similarity Metric (cSM) for the evaluation and optimization of GANgenerated skin lesion images. The proposed metric provides a more informative assessment of synthetic image quality than conventional evaluation approaches. In addition, a pseudo-label–based evaluation strategy is introduced to assess the realism and utility of synthetic melanoma images, demonstrating their potential to enhance classification performance when incorporated into training datasets. The third study investigates whether clinically relevant attributes,such as patient age, gender, anatomical lesion location, and lesion class are implicitly encoded within dermoscopic images. A multi-task learning framework is employed to predict these attributes simultaneously from a single image. The findings suggest that while certain attributes can be inferred with moderate accuracy, others remain challenging, revealing both limitations in current deep learning representations and opportunities for future model and dataset improvements. The final part of the thesis addresses skin lesion segmentation and its influence on melanoma segmentation and classification. A novel segmentation-based framework, DESeg SkinNet,is developed for segmentation efficiency and accuracy, and to evaluate the impact of background removal on classification performance. This study highlights the importance of efficient lesion segmentation and optimized preprocessing pipelines in improving downstream classification results. Overall, this thesis contributes to the field of medical image analysis by providing a comprehensive investigation into the role of data curation, synthetic data generation, and segmentation in deep learning–based skin cancer diagnosis. The findings underscore the importance of high quality datasets and tailored preprocessing strategies for achieving reliable performance. The proposed methods demonstrate strong potential for improving early melanoma detection and offer valuable insights for the development of more robust and clinically applicable diagnostic systems.