Deep Learning-Based Automatic Diagnosis of Diseases

Abstract

Skin cancer is a growing global health concern and remains a major global health challenge due to its aggressive nature and the complexity of early diagnosis. Early diagnosis is the secret to improved patient outcomes, especially when certain subtypes, such as nodular melanoma or atypical nevi, continue to pose classification difficulties even for experienced clinicians. The increasing worldwide incidence of skin cancer highlights the pressing need for accurate, accessible, and early diagnostic tools that can support clinical decision-making and improve patient outcomes. In recent years, artificial intelligence (and more specifically, deep learning) has emerged as a powerful ally in medical imaging. Within dermatology, deep convolutional neural networks (CNNs) have shown great promise in assisting with skin lesion analysis and improving diagnostic performance. The motivation behind this thesis stems from the goal of enhancing public health through the development of intelligent, efficient, and cost-effective classification tools that harness technological innovation. Contributions Synthesis Our work is aimed to improve the field of computer-aided dermatological diagnosis by exploring and optimizing deep learning-based classification systems for skin cancer, with specific focus on binary discrimination between malignant and benign lesions. Through a multi￾faceted approach grounded in real-world imaging challenges and modern neural network architectures, we addressed critical limitations in dataset quality, class imbalance, captured images configuration bias, and diagnostic variability. We took advantage of the SLICE-3D dataset, a recent dermoscopic image database revealed in the ISIC 2024 challenge. Among the core challenges inherent in this dataset; and indeed most clinical imaging databases; was the extreme class imbalance with malignant samples being an incredibly small minority. To offset this, we employed a DCGAN-based image synthesis pipeline that successfully augmented malignant cases to balance the dataset. The synthetic images, when validated by both visual quality and downstream classification performance, were found to be an effective method for improving data diversity without compromising quality. Yet another notable aspect of this thesis was the particular focus devoted to images’ lighting modality, which is usually given short shrift but has quite a high degree of influence in dermoscopic imaging especially after researching and finding only few AI studies in this domain, they were all purely medical researchers conducted with dermatologists and not AI models. So, with the assistance of metadata fields such as tbp_tile_type, we investigated this relatively unexplored area and separated and tested independently three lighting conditions: non-polarized white light (3D:white), Cross-polarized light (3D:XP), and an intermediate set Y.Z. MECIFI 73 General conclusion combination. This allowed us to explore systematically the effect of light on automated diagnostic model performance; a study that, as far as we know, had not been previously investigated with such extent of detailed comparison in Deep learning filed. Four pre-trained convolutional neural networks, DenseNet121, ResNet50, MobileNetV2, and EfficientNet-B0, were tested to perform binary classification on the preprocessed datasets. The experimental results consistently demonstrate that dermoscopic lighting conditions have a decisive influence on CNN-based skin cancer classification performance. Across all evaluated architectures, non-polarized (white-light) dermoscopy yielded the highest Recall, Specificity, and AUC, confirming its superiority for detecting malignant lesions while maintaining reliable benign exclusion. This finding is clinically significant, as high Recall is essential for screening￾oriented diagnostic systems. Among the tested models, EfficientNet-B0 achieved the best overall balance between accuracy and discriminative performance, particularly under white-light illumination, while MobileNetV2 and ResNet50 exhibited exceptionally high Recall and Specificity, making them well suited for sensitivity-critical clinical scenarios. These results align with dermatological knowledge, as white-light dermoscopy enhances superficial epidermal and pigmentation￾related structures that are highly informative for malignancy discrimination. The key contributions of the thesis are: • Successful implementation of DCGAN to address class imbalance in our clinical dataset and crafting a transfer learning model that achieved a very high and impressive level of accuracy. • Benchmarking dermoscopic lighting conditions (XP vs. White Light) on DCNN performance. • Methodologically rigorous comparison of multiple architectures across balanced datasets, supported with good metrics. • Demonstrating that metadata-aware experimentation can produce more clinically valuable models. Overall, this work emphasizes the power of deep learning especially when combined with synthetic data generation and imaging metadata to improve early detection of skin cancer. Future Research Directions and Perspectives The findings and contributions of this thesis open several promising avenues for further research aimed at advancing deep learning methodologies for reliable and clinically applicable skin cancer detection: • Exploration of More Advanced Generative Models for Data Augmentation: While this thesis investigated DCGAN for synthetic image generation, future work Y.Z. MECIFI 74 General conclusion could explore other GANs like StyleGAN or diffusion-based generative models and hybrid architectures. These approaches may yield higher-fidelity lesion synthesis, more diverse feature representation, and improved generalization when integrated into training pipelines. • Multi-Modal and Cross-Domain Learning: Extending the analysis beyond dermoscopy images by incorporating additional modalities such as clinical photographs, histopathology, or metadata (e.g., patient age, lesion location) could significantly enhance diagnostic robustness. Research into multi￾modal fusion strategies and domain adaptation techniques would support broader generalizability across institutions and imaging devices. • Systematic Investigation of Lighting Modalities: This thesis highlighted the impact of lighting conditions (white light, cross-polarized, mixed) on model performance. Future work should expand this study by including standardized multi-center datasets, quantifying robustness under diverse acquisition protocols, and developing adaptive models capable of automatically normalizing or correcting lighting variations. • Explainability and Clinical Trustworthiness of Models: Future studies should place stronger emphasis on explainable AI (XAI) methods tailored to dermatology. Designing interpretable saliency maps, lesion-specific feature attribution, or prototype-based reasoning can help bridge the gap between model predictions and dermatologists’ diagnostic reasoning, thus enhancing trust in clinical settings. • Transferability and Federated Learning Across Institutions: To overcome data scarcity and privacy concerns, research into federated learning and privacy-preserving training approaches for skin lesion datasets is recommended. This would allow collaboration across hospitals without direct data sharing, while ensuring that models remain robust and clinically validated across diverse populations. • Robustness under Data Shifts and Longitudinal Generalization: Future investigations should examine how models behave under domain shifts such as population diversity, new lesion subtypes, or evolving imaging technologies. Research into continual learning and model adaptation will be crucial for ensuring that AI systems maintain high diagnostic accuracy in dynamic real-world clinical environments. Overall, this work emphasizes the power of deep learning, especially when combined with synthetic data generation and imaging metadata, to improve early detection of skin cancer. As Y.Z. MECIFI 75 List of Published Work encouraging as the results are, there is potential future work that can enhance diagnostic aid systems.