Volume 11 Issue 1

Enhanced SVM-Based Model for Skin Cancer Detection Using Dermoscopic Images

Sotonwa

Abstract

Introduction: Skin cancer is still a pervasive worldwide health concern, amid its increasing cases which continues to pose a burden on global health challenge. Melanoma, a very ferocious type of skin cancer, requires timely detection owing to its tendency to spread quickly throughout the body. Timely diagnosis markedly increases survival outcomes, but maintaining reliability in clinical evaluation remains an intractable problem despite advances in AI. Dermoscopy, a widely used non-invasive imaging technique, strengthens the visualization of deeper skin layers and significantly aids in recognizing abnormal lesions. However, interpreting these images is not straightforward, it requires years of expertise. Recent breakthroughs in artificial intelligence (AI) have created valuable prospects to improve skin cancer detection. While deep learning (DL) and machine learning (ML) models have each exhibited effectiveness in analyzing dermoscopic images, individually, these methods are limited by challenges. By blending the advantages of different algorithms, hybrid AI models have emerged as a powerful solution, which boost feature extraction, accuracy, and reliability in ways single models often cannot. By integrating synergies, such models support dermatologists with faster, more consistent diagnostic outputs, allowing skin cancers such as melanoma to be recognized sooner and treated more effectively.
Aim: In this work, we propose a SVM-based diagnostic approach that aims to improve the classification accuracy of lesions.
Method: A total of 5,000 dermoscopic images depicting nine different skin lesion types were utilized. For preprocessing, images were standardized using lesion segmentation, color normalization, resizing, and augmentation. EfficientNet-B7 (pre-trained on ImageNet) served as a feature extractor.
Results: The SVM-Based model approach demonstrated strong performance, achieving an accuracy of 88.00% and an AUC-ROC of 88.63%. Primarily, the model managed to deliver an outstanding malignant sensitivity of 99.00%. This means that the system detects almost every true case of cancer.
Conclusion: This research progressive advancement in the area of malignant metrics compared to the previous hybrid model (F1-score 93.00% and 91.00%), which further confirms its reliability and trust. This is achieved by combining the powerful representation features of EfficientNet-B7 with SVM classification, the model focused on detecting most malignancies and obtained low false negatives.

Keywords

Full Text

Download

References