The Clinical Motivation Melanoma is among the most rapidly increasing cancers worldwide, but it can be cured by simple excision if detected early. Dermoscopy - a non-invasive imaging technique using optical magnification and polarized light to visualize subsurface skin structures - improves diagnostic accuracy over naked-eye examination, but requires significant training and introduces diagnostic subjectivity between observers.
Computerized Analysis as a Solution Automated dermoscopy image analysis can serve as an objective, reproducible second opinion, help non-specialist clinicians make more accurate biopsy decisions, and assist in screening large numbers of patients. However, the field at the time of this 2007 publication lacked a rigorous systematic methodology with transparent feature selection and validated classification.
What This Paper Contributes Celebi and colleagues present a complete end-to-end methodology covering every step from raw image to diagnosis: automatic lesion border detection, extraction of shape, color, and texture features, feature selection using multiple ranking algorithms, support vector machine classification with class imbalance correction, and rigorous cross-validation using Monte Carlo simulation.
Key Results On a dataset of 564 dermoscopy images (88 melanoma, 476 benign), the optimized pipeline achieved 92.34% specificity and 93.33% sensitivity for melanoma detection - performance competitive with and in some comparisons exceeding trained dermatologists working without dermoscopy at that time.
Why Border Detection is Critical The first step in any automated dermoscopy analysis is separating the lesion from background skin. Border accuracy matters for two reasons: the shape of the border itself is diagnostically important (asymmetry, irregularity), and correct border localization is required for computing clinically meaningful color and texture features from the right image regions.
The JSEG-Based Detection Method The border detection algorithm was based on JSEG (J-measure Segmentation), a color image segmentation method. The pipeline included three phases: preprocessing (color median filtering, variance-based color quantization, rough lesion localization via Otsu thresholding); segmentation (computing J-images that mark potential boundary locations, region growing, and region merging by color similarity in the CIE L*u*v* perceptually uniform color space); and postprocessing (removing background skin regions, eliminating isolated segments, and merging remaining regions).
Performance and Failure Cases The border detection was applied to 596 images. Thirty-two were excluded for unsatisfactory results, leaving 564 for the main analysis. Failures occurred primarily in two scenarios: lesions with a very smooth gradient transition between lesion and skin (indistinct borders), and lesions with regression areas (scar-like depigmentation), both of which produce color patterns that confuse the segmentation algorithm.
Dataset Composition The final 564-image set included 88 melanoma and 476 benign lesions, representing substantial class imbalance (a 5.4:1 benign-to-melanoma ratio). Benign lesions included Clark nevi (the most common), Reed/Spitz nevi, seborrheic keratoses, compound nevi, blue nevi, dermatofibromas, vascular lesions, and lentigines. Diagnosis was confirmed histopathologically in cases with significant melanoma risk, otherwise by long-term clinical follow-up.
Shape Features from the Border Eleven shape features were extracted from the lesion border: area (computed via bit quads for accuracy on rough borders), aspect ratio, two asymmetry measures (computed by folding the lesion along major and minor axes), compactness (ratio of equivalent to maximum diameter), maximum lesion diameter, eccentricity, solidity (ratio of lesion area to convex hull area - measuring border irregularity), equivalent diameter, rectangularity, and elongation. Border length-dependent features like perimeter and circularity were excluded due to estimation inaccuracies.
Clinically Significant Color Regions Rather than extracting color features from the entire lesion uniformly, the image was divided into concentric regions using Euclidean distance transform, creating zones progressively more distant from the lesion center. This captures the clinically important distinction between peripheral and central lesion color patterns - a feature used by dermatologists when applying diagnostic criteria like ABCDE and dermoscopic pattern analysis.
Color Features For each of the concentric regions and across multiple color spaces (RGB, HSV, CIE L*a*b*, opponent color), statistical features were extracted: mean, standard deviation, skewness, kurtosis, and percentile values. Additionally, color histogram features and percentages of dominant colors were computed. The multicolor-space approach was motivated by evidence that different color spaces capture different clinically relevant color information in pigmented lesions.
Texture Features Texture was characterized using gray-level co-occurrence matrix (GLCM) features computed at multiple orientations and distances, capturing spatial dependencies between pixel intensities. Gabor filter responses at multiple frequencies and orientations provided additional texture information. These texture features encode properties such as lesion smoothness, coarseness, and structural regularity that relate to pigment network patterns visible in dermoscopy.
Feature Selection Framework Rather than using all extracted features (which risk overfitting and increased computation), the paper employed an optimization framework that ranked features using multiple independent algorithms: Fisher's discriminant ratio, t-test, entropy-based measures, and Relief-F. The optimal subset size was determined by maximizing the area under the ROC curve from a support vector machine classifier, ensuring the selection criterion aligned with the ultimate diagnostic goal.
Why Multiple Feature Selection Methods? Different feature selection algorithms make different assumptions about class separability. Using multiple methods and examining their agreement provides a more robust estimate of feature importance and reduces the risk of selecting features that perform well only under the assumptions of one specific algorithm.
Handling Class Imbalance With nearly 5.4 benign lesions for every melanoma, standard classifiers tend to optimize overall accuracy by mostly predicting 'benign,' achieving high accuracy through insensitivity to the rare class. The paper evaluated several sampling strategies to address this imbalance: random oversampling of the minority class, random undersampling of the majority class, and SMOTE (Synthetic Minority Over-sampling Technique), which generates synthetic minority class examples.
Support Vector Machine Classification Classification used a support vector machine (SVM) with radial basis function (RBF) kernel, a well-validated method for high-dimensional medical classification problems. The SVM finds the maximum-margin decision boundary that separates the two classes in the transformed feature space. Generalization error was estimated using Monte Carlo cross-validation with 100 random train/test splits, providing more reliable performance estimates than simple k-fold cross-validation on a small dataset.
Best Performance Achieved The optimized pipeline - combining the best feature subset with SMOTE oversampling and SVM classification - achieved 92.34% specificity and 93.33% sensitivity. This means that 92.34% of benign lesions were correctly identified as benign (avoiding unnecessary biopsy) and 93.33% of melanomas were correctly identified as malignant (avoiding missed diagnoses).
Most Informative Features Feature selection analysis identified that color and texture features contributed most to classification performance, with shape features adding complementary but less dominant information. Specifically, features from the peripheral lesion regions (outer concentric zones) were more informative than central regions - consistent with the clinical observation that peripheral asymmetry and irregular color distribution are key melanoma indicators.
Impact of Imbalance Correction Without class imbalance correction, classifiers achieved high specificity (correctly identifying benign lesions) but poor sensitivity (missing melanomas). SMOTE substantially improved sensitivity with only a modest reduction in specificity, demonstrating that addressing class imbalance is essential for clinically useful melanoma detection rather than just high overall accuracy.
Comparison to Prior Literature The paper systematically compared results to existing literature (Table 1 in the original), noting that the achieved performance was competitive with or superior to prior automated systems and to unaided clinical diagnosis, while using a substantially larger and more diverse dataset than most prior studies, which used images from single sources.
Limitations of the Feature Engineering Approach The Celebi 2007 pipeline exemplifies classical machine learning: human experts define features (shape, color, texture), algorithms select the best features, and a classifier is trained. This approach requires substantial domain expertise to design features and may miss patterns that are diagnostically important but not captured by predefined feature types.
The Deep Learning Revolution By the mid-2010s, convolutional neural networks trained end-to-end on dermoscopy images began to supersede feature-engineering approaches, as exemplified by the Esteva et al. 2017 Nature paper and the ISIC challenge results. Deep learning learns feature representations directly from the training images, often capturing patterns invisible to human-designed feature descriptors.
Continuing Relevance of Classical Methods Feature-engineered approaches remain valuable in settings with limited training data, where interpretable features are required for regulatory approval or clinician trust, or where computational resources are constrained. The Celebi 2007 methodology also established the evaluation standards (ROC-AUC, Monte Carlo cross-validation, explicit handling of class imbalance) that later deep learning papers in dermoscopy were expected to meet.
Hybrid Approaches Modern systems often combine deep learning feature extraction with interpretable feature analysis - using neural networks to extract high-level features while maintaining the ability to explain predictions in terms of clinically meaningful properties like asymmetry, border irregularity, and color variation. This hybridization aims to preserve the diagnostic transparency of feature-engineering approaches while leveraging the representation power of deep learning.