Two-stage pipeline approach This study proposes a hybrid architecture for automated melanoma classification from dermoscopy images. Rather than training a single end-to-end deep learning classifier, the authors separate the problem into two sequential stages: a deep learning U-Net for precise lesion segmentation, followed by handcrafted feature extraction and classical machine learning classifiers applied to the segmented region.
Why segmentation matters Accurate delineation of the lesion boundary is a prerequisite for reliable feature extraction. Without segmentation, features are computed across the entire image including irrelevant skin, hair, and background artifacts. By segmenting first, the pipeline focuses all subsequent analysis on the diagnostically relevant lesion tissue.
Key results The U-Net segmentation achieved a Dice coefficient of 77.5%, while the subsequent SVM classifier trained on features from segmented lesions achieved 85.19% overall accuracy for melanoma classification. The study compared SVM against random forest, k-nearest neighbors, and naive Bayes classifiers to identify the best performing algorithm.
U-Net encoder-decoder architecture The segmentation model uses U-Net, a convolutional neural network designed specifically for biomedical image segmentation. U-Net's encoder progressively extracts high-level semantic features while its decoder reconstructs spatial resolution through upsampling, with skip connections linking corresponding encoder and decoder layers to preserve fine-grained boundary detail.
Dermoscopy-specific challenges Skin lesion segmentation is complicated by low contrast at lesion borders, presence of hair, skin markings, illumination gradients, and highly variable lesion shapes. U-Net's pixel-wise classification approach handles these challenges by learning context at multiple scales simultaneously, rather than relying on simple thresholding methods.
Segmentation performance The trained U-Net achieved a Dice similarity coefficient of 77.5% on the test set. While this falls below state-of-the-art benchmarks, it was sufficient to isolate lesion regions for subsequent feature extraction more reliably than traditional image processing segmentation methods used in prior work.
Feature extraction strategy After segmentation, four complementary texture descriptors were extracted from the isolated lesion region: Local Binary Patterns (LBP) for local texture microstructure, Edge Histogram (EH) for directional edge distribution, Histogram of Oriented Gradients (HOG) for shape and gradient patterns, and Gabor filter responses for spatial frequency and orientation content.
Complementary descriptors Each descriptor captures a distinct aspect of lesion appearance. LBP characterizes whether melanoma lesions have irregular microstructures versus smooth benign nevus patterns. HOG captures the irregular border morphology of melanoma - the dermoscopic 'ugly duckling' criteria. Gabor features identify the atypical pigmentation networks characteristic of malignant lesions.
Feature concatenation Features from all four descriptors were concatenated into a single feature vector per lesion, combining local texture, edge, shape, and frequency information into a comprehensive representation. This multi-descriptor approach aims to capture complementary diagnostic signals that no single feature type alone could provide.
Classifier comparison Four classifiers were evaluated on the extracted feature vectors: support vector machine, random forest, k-nearest neighbors, and naive Bayes. The SVM with radial basis function kernel achieved the highest accuracy at 85.19%, outperforming random forest, k-nearest neighbors, and naive Bayes classifiers on this dataset.
Why SVM excels here SVMs find the optimal hyperplane separating classes in high-dimensional feature spaces, making them well-suited for the concatenated multi-descriptor feature vectors used in this pipeline. The kernel trick allows non-linear separation in feature space, capturing complex interactions between texture, edge, and shape features that linear classifiers would miss.
Performance interpretation An 85.19% accuracy represents meaningful improvement over visual-only assessment by general practitioners, though expert dermatologists in controlled studies achieve higher accuracy. The pipeline's strength is its potential for deployment in resource-limited settings where specialist access is unavailable.
Interpretability advantage The hybrid approach - segmentation followed by handcrafted feature extraction and classical classification - provides clearer diagnostic transparency than end-to-end CNNs. Feature values can be inspected directly, and each texture descriptor maps onto known dermoscopic criteria such as irregular pigmentation network or atypical vascular patterns.
Performance ceiling End-to-end CNN classifiers such as EfficientNet or ResNet trained directly on dermoscopy images consistently outperform handcrafted feature pipelines on large benchmarks like ISIC 2018 and 2019. The 85.19% accuracy achieved here is competitive for a hybrid approach but below the 90%+ accuracy of state-of-the-art end-to-end networks.
Computational accessibility Handcrafted feature extraction is computationally inexpensive compared to training deep CNN classifiers on large labeled datasets. For deployment in low-resource clinical settings with limited GPU infrastructure or small local training datasets, hybrid pipelines may offer a practical alternative to deep learning classifiers that require large labeled training collections.
Segmentation quality as the bottleneck With a Dice coefficient of 77.5%, segmentation errors propagate into the feature extraction stage and degrade classification. Improving segmentation using more training data, data augmentation, or attention-enhanced U-Net variants would directly improve downstream classification accuracy without changing the feature extraction or classification pipeline.
Multi-class skin lesion classification This study focused on binary melanoma versus non-melanoma classification. Extending the pipeline to multi-class differentiation among the eight major ISIC skin lesion categories would require redesigning the classifier head and collecting balanced multi-class training data, but would significantly increase clinical utility.
Integration of deep feature learning A promising extension is replacing or supplementing handcrafted descriptors with CNN-extracted features from the segmented lesion region. Deep features from pre-trained networks like VGG16 or ResNet applied to segmented lesion crops could substantially improve accuracy while retaining the localization benefit that the segmentation step provides.