EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Deep Learning-Based Segmentation and Seven-Class Classification of Dermoscopic Skin Lesions: Model Development and Internal Test-Set Evaluation

Orkhan Majidov, Arin Özbaltacı

Computer and Decision Making: An International Journal · 03.11.2026 · 3 (6S) · 3156-3170

DOI: 10.1922/comdem.v3i6s.19060 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Computer and Decision Making: An International Journal, online erschienen am 03.11.2026. Verfasst von Orkhan Majidov und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

The early and accurate diagnosis of skin cancer, particularly melanoma, is important. Image analysis is one of the possible approaches that can be automated in skin lesion assessment and is supported by deep learning. The aim of this study was to develop and evaluate a deep learning approach for automated skin lesion segmentation and multiclass classification of dermoscopic images. Segmentation of the lesions was performed using a U-Net architecture with a ResNet-34 encoder, and classification of seven classes of lesions was performed using DenseNet-121. For classification, we used the HAM10000 dataset, and for segmentation, the ISIC 2016 dataset was used. The images were resized and augmented through rotation, horizontal flipping, vertical flipping, and brightness adjustment. Transfer learning was used with ImageNet-pretrained weights. The segmentation performance was measured by pixel accuracy, Jaccard score, and Dice score. Classification performance was measured by sensitivity and specificity. For the seven-class classification task, sensitivity and specificity were computed for each diagnostic class using a one-versus-rest approach, and the reported overall values were calculated using micro-averaging across the seven one-versus-rest class comparisons. The segmentation model obtained a pixel accuracy of 0.91, Jaccard score of 0.73, and Dice score of 0.89. Overall multiclass sensitivity and specificity were 0.78 and 0.97, respectively, using micro-averaging across the seven one-versus-rest class comparisons. However, there was some misclassification of melanocytic nevi as melanoma cases. The confusion matrix showed high recognition of melanocytic nevi, while discrimination between visually similar lesion categories remained difficult. The proposed framework is feasible for automated segmentation of skin lesions and classification into seven classes. Overall segmentation performance was good, but class-specific performance and melanoma misclassification indicate limitations in sensitivity. The framework needs to be further optimized, class balanced, and independently validated before potential clinical use.

Zitieren

APA

Majidov, O., & Özbaltacı, A. (2026). Deep Learning-Based Segmentation and Seven-Class Classification of Dermoscopic Skin Lesions: Model Development and Internal Test-Set Evaluation. Computer and Decision Making: An International Journal, 3(6S), 3156-3170. https://doi.org/10.1922/comdem.v3i6s.19060

Vancouver

Majidov O, Özbaltacı A. Deep Learning-Based Segmentation and Seven-Class Classification of Dermoscopic Skin Lesions: Model Development and Internal Test-Set Evaluation. Computer and Decision Making: An International Journal. 2026;3(6S):3156-3170. doi:10.1922/comdem.v3i6s.19060

BibTeX
@article{Majidov2026,
  author = {Majidov, Orkhan and Özbaltacı, Arin},
  title = {Deep Learning-Based Segmentation and Seven-Class Classification of Dermoscopic Skin Lesions: Model Development and Internal Test-Set Evaluation},
  journal = {Computer and Decision Making: An International Journal},
  year = {2026},
  volume = {3},
  number = {6S},
  pages = {3156--3170},
  doi = {10.1922/comdem.v3i6s.19060}
}
Kostenlos starten →