Introduction: Skin cancer incidence is increasing globally, making early and accurate diagnosis essential. Dermoscopy improves detection but is limited by clinician variability. Automated segmentation using deep learning offers a promising alternative, though challenges such as image variability and artifacts remain.
Aims: To develop and evaluate an improved UNet++-based model for accurate skin lesion segmentation and classification in dermoscopic images.
Materials and Methods: The study utilised the ISIC 2018 dataset comprising 2,694 dermoscopic images with corresponding expert-annotated segmentation masks. A UNet++ architecture with an ImageNet-pretrained EfficientNet-B5 encoder was implemented. Data augmentation techniques, including geometric and photometric transformations, were applied to improve generalisation. A composite loss function combining Dice loss (50%), binary cross-entropy (30%), and focal loss (20%) was used to address class imbalance and improve boundary delineation. Model performance was evaluated using sensitivity, specificity, accuracy, Dice coefficient (DC), Intersection over Union (IoU), F1-score, and area under the curve (AUC).
Results: The model achieved classification sensitivity of 0.8900, specificity of 0.9200, accuracy of 0.9500, F1-score of 0.8870, and AUC of 0.9278. For segmentation, it achieved a Dice coefficient of 0.8870 and an IoU of 0.8200. The model outperformed U-Net and showed consistent improvements over O-Net across key metrics. Conclusion: The proposed UNet++ framework improves segmentation accuracy and classification performance, demonstrating strong potential for clinical application in automated skin cancer diagnosis.