Loading...
Precise Skin Cancer Segmentation: Unveiling Insights through Total Variation and Multiresolution Analysis
Muthanna, Lina Hassan, Rana Muqaybil, Yara
Muthanna, Lina
Hassan, Rana
Muqaybil, Yara
Citations
Altmetric:
Type
Supervisor
Subject
Date
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
This study investigates advanced image processing techniques for skin lesion classifica- tion using the HAM10000 dermatoscopic dataset. A comprehensive approach is taken, addressing noise removal, image enhancement, artifact elimination, segmentation, feature extraction, and classification. Various noise reduction techniques, including Wiener filter- ing and Non-Local Means (NLM), were evaluated, with the Wiener filter performing best in terms of signal-to-noise ratio (SNR) and peak signal-to-noise ratio (PSNR). Classical and adaptive image enhancement methods were compared, with Contrast-Limited Adap- tive Histogram Equalization (CLAHE) and Multi-Scale Retinex (MSR) showing e↵ective contrast improvement without introducing noise. Hair artifact removal was achieved us- ing blackhat morphological operations combined with Telea’s inpainting method, demon- strating high accuracy in preserving critical diagnostic features. For segmentation, the Total Variation Chan-Vese model was applied to skin lesions, demonstrating strong performance across di↵erent lesion types as evaluated by the Jac- card Index. Feature extraction focused on frequency, shape, texture, and wavelet features, with results showing that combining these feature sets and applying dimensionality re- duction enhances classification accuracy. Machine learning algorithms, including Random Forest (RF), Support Vector Machine (SVM), XGBoost, and Artificial Neural Networks (ANN), were utilized for lesion classification, with a Stacking Ensemble (SEns) model integrating RF and XGBoost achieving the highest accuracy of 92% on the test dataset. The study highlights the e↵ectiveness of ensemble learning and robust preprocessing in overcoming challenges posed by imbalanced datasets, paving the way for reliable auto- mated diagnostic systems for skin lesion analysis.rite your abstract text here.