부산대학교 도서관

This research focuses on Melanoma skin cancer, an often-occurring condition originating in melanocytes responsible for melanin production. The study employs a deep hybrid learning model to effectively discriminate and categorize skin cancers, utilizing a dataset comprising benign and malignant classes. To address the dataset's imbalance, particularly the scarcity of images depicting malignant lesions, augmentation techniques are applied. Additionally, Contrast Limited Adaptive Histogram Equalization (CLAHE) is utilized to enhance image clarity. Lesion segmentation is conducted using a neural network-based ensemble model, combining segmentation algorithms from Fully Convolutional Network (FCN), SegNet, and U-Net. This results in a binary representation of the skin and lesion, with lesions in white and the skin in black. Subsequently, these binary images undergo further classification using various pre-trained models such as Inception ResNet V2, Inception V3, ResNet 50, Densenet, and CNN. The best-performing pre-trained model undergoes fine-tuning for enhanced classification performance. To augment the classification model's effectiveness, a combination of deep learning (DL) and machine learning (ML) is applied. DL models are utilized to extract features, while classification is executed using Support Vector Machines (SVM). This computer-aided tool is designed to expedite disease diagnosis for medical professionals compared to conventional methods, resulting in a notable 4% improvement in the proposed method's performance.