Sürdürülebilir Kalkınma Amaçları
Yayınlar
- BildiriYalnız künye2025Improving Skin Lesion Classification with Pretrained Deep Learning Models
Almasani, Heba; Ture, Hayati
Skin cancer is the most common type of cancer. Cancer can be defined as a disease characterized by the uncontrolled division and spread of cells. As with other types of cancer, early diagnosis and treatment play a crucial role in combating skin cancer. Since diagnosis is made through visual examination, factors such as the level of expertise and image artifacts negatively affect diagnostic accuracy and slow down the process.Computer-aided applications used to analyze such images provide faster and more accurate results. Additionally, these methods eliminate factors that reduce diagnostic accuracy during the classification and segmentation of skin lesions.This study classifies skin cancer to benign or malignant using deep learning and machine learning methods. In particular, Convolutional Neural Networks (CNNs) have shown promising results in medical image classification. The study evaluates six pre-trained CNN models - VGG16, EfficientNetB7, EfficientNetV2L, DenseNet, ConvNeXtBase and MobileNet - for binary skin lesion classification using the ISIC 2019 dataset. The study first preprocesses and balances the data to enhance performance. The results indicate that the ConvNeXtBase model achieved the highest success with accuracy rate of 84.63% after balancing the data that we have.
- MakaleAçık Erişim2025A deep learning approach powered by GhostNet for skin cancer classification; [GhostNet tabanlı derin öğrenme yaklaşımı ile cilt kanseri sınıflandırması]
Almasani, Heba; Ture, Hayati
Skin cancer, especially melanoma, continues to cause a disproportionate percentage of cancer-related deaths despite its high curability when detected early. This work presents an efficient deep-learning system aimed at the binary classification of dermoscopic images as malignant or benign using the GhostNet family of convolutional neural networks (CNNs). The publicly available ISIC-2019 dataset was used for training its extreme class imbalance was corrected using a hybrid Synthetic Minority Over-sampling Technique coupled with Edited Nearest Neighbours (SMOTE-ENN). End-to-end fine-tuning of three variants of GhostNet (V1, V2, V3) was performed with preprocessing, augmentation, and hyper-parameters kept constant to ensure a fair model-to-model comparison. Evaluation metrics used were accuracy, precision, recall, F1-score, and the area under the receiver-operating-characteristic curve (AUC). GhostNet V2, augmented with depth-wise attention, gave the strongest results: 95% accuracy, 94% malignant class recall, F1-score of 94.3%, and an AUC of 0.99. GhostNetV2 performed better than both the baseline V1 and V3 without having a parameter count that would prevent real-time inference on mobile hardware. These results of Accuracy, sensitivity, F1-score, AUC and recall show that, when combined with targeted imbalance correction, efficient architectures such as GhostNet are capable of dermatologist-level sensitivity without the computational requirements of heavier models, and thus are feasible for point-of-care or resource-constrained settings. It is complementary to our previous work with other CNNs, allowing model-to-model comparison directly.