The segmentation problem in dermatology: Accurate delineation of skin lesion boundaries is a prerequisite for quantitative diagnosis using the ABCD rule - Asymmetry, Borders, Colors, and Dermatoscopic Structures. Manual contour tracing by dermatologists is time-consuming and variable. Automated segmentation methods must separate the lesion region from surrounding normal skin, but performance depends critically on how well the chosen color representation distinguishes the lesion from healthy skin.
Limitations of standard color spaces: Commonly used color transformations like CIE L*a*b*, L*u*v*, and RGB are image-independent - they were designed as general-purpose perceptual color representations, not for any specific type of lesion or imaging system. The optimal color channel for separating a lesion from normal skin varies across lesion types (melanoma, dysplastic nevus, psoriasis) and across different imaging devices. A fixed color space will fail in cases where the lesion-skin contrast is low in all standard channels.
The IHP approach: Independent Histogram Pursuit (IHP) is an unsupervised algorithm that finds image-specific linear combinations of spectral bands that maximize the bimodal histogram separation between lesion and normal skin pixels. Unlike PCA (which maximizes variance) or fixed color transforms, IHP directly optimizes for segmentation - finding the projection direction that creates the clearest two-group separation in pixel intensities.
Key result: On five diverse dermatological datasets including melanoma, nevi, and psoriasis images, IHP achieved boundary detection precision close to 97% after standard morphological postprocessing. The algorithm also works with multispectral images (beyond standard RGB), making it applicable to new imaging modalities being developed for dermatological diagnosis.
Core concept - bimodal pixel distribution: A typical dermatological image consists of two large pixel classes: lesion and normal skin. If the right projection direction is found, the histogram of pixel values projected onto that direction will be bimodal - two well-separated peaks corresponding to the two tissue types. IHP's objective is to find this optimal projection direction by maximizing the concavity area between the two histogram modes.
Optimization formulation: For an N-band image, IHP seeks N orthogonal vectors {u1, ..., uN} computed sequentially. The first component u1 maximizes the area of the concavity between the two modes of the histogram of whitened pixel values projected onto the direction z. Whitening (removing correlations between bands) is a prerequisite for making the estimated components statistically independent. The optimization is performed using a Genetic Algorithm (GA) with population size 100 and 200 iterations.
Subsequent components for lesion substructure analysis: After finding the primary lesion-skin separation direction u1, the remaining N-1 components are computed under the orthogonality constraint (using Gram-Schmidt orthogonalization). These subsequent components enhance less obvious structures within the image - such as borders, vessels, hair, or pigmentation patterns. For psoriasis images, the second component revealed the characteristic red inflammatory zone and scales; for melanoma, it enhanced the lesion border.
Segmentation execution: Once the first IHP component is computed, pixels are projected onto this direction to produce a 1D grayscale representation. Standard k-means clustering (k=2) is then applied to this projected image to segment lesion from normal skin. Morphological postprocessing (closing, opening, hole filling, and selection of the largest blob) further refines the boundary, pushing precision from the raw IHP segmentation to 97%.
Primary melanoma dataset (Chen et al.): The first evaluation used 100 RGB melanoma dermoscopy images (70 malignant melanomas, 30 dysplastic nevi). IHP outperformed all tested standard color channels: RGB (R, G, B), CIE L*a*b* (L, a, b), L*u*v* (L, u, v), and XYZ - as well as PCA components. The first IHP component achieved the highest segmentation accuracy against dermatologist-drawn reference contours. Compared with JSEG, watershed, and GVF snake algorithms - the best methods from the previous literature on this dataset - IHP with postprocessing showed superior or comparable boundary detection.
Ganster melanoma dataset (150 images): A second evaluation on 150 melanoma images compared IHP against Ganster's fusion algorithm, which combined five different segmentation methods (global thresholding, three adaptive thresholds, and 3D color clustering). Despite this complexity, IHP achieved comparable performance to the fusion algorithm while being simpler - just a single linear band combination rather than five separate techniques. Critically, Ganster's algorithm used color spaces empirically chosen for this specific dataset, while IHP was fully automatic.
Hintz-Madsen melanoma dataset (57 images): On a third melanoma dataset, IHP correctly segmented all 57 images without any preprocessing other than hair removal. This dataset lacked manual segmentations, so visual evaluation confirmed that the detected boundary always matched the lesion boundary. The algorithm's success without any preprocessing demonstrated robustness to natural image variation.
Psoriasis datasets (RGB and multispectral): Two psoriasis datasets - one RGB (35 images) and one 9-band multispectral spanning ultraviolet to near-infrared (9 images) - confirmed that IHP generalizes beyond melanoma and beyond RGB imaging. For the multispectral images, IHP discovered that the optimal lesion-enhancing projection used all nine spectral bands, demonstrating that it can leverage information unavailable to standard color-space methods.
Image-independence advantage: The most significant clinical advantage of IHP is its automatic adaptation to each specific image without requiring parameter tuning or prior training. A system that works reliably across different dermoscopes, lighting conditions, and lesion types can be deployed in diverse clinical environments - from specialized dermoscopy laboratories to general practitioner offices with standard cameras.
Foundation for ABCD analysis: Lesion segmentation is the necessary first step in any automated ABCD rule implementation. Once the lesion boundary is accurately delineated by IHP, downstream algorithms can compute asymmetry scores, border irregularity indices, color heterogeneity measures, and dermatoscopic structure patterns. Errors in segmentation propagate into all subsequent ABCD features, making the 97% boundary detection precision of IHP directly clinically significant.
Multispectral imaging potential: Standard dermoscopes capture RGB images, but next-generation devices acquire multispectral images from UV to near-infrared. These additional wavelengths penetrate different skin depths and reveal subsurface structures invisible in RGB. IHP's band-independent formulation means it can directly leverage multispectral images as dermatological imaging technology advances, without algorithmic modification.
Compatibility with classification algorithms: IHP outputs a segmented lesion region that can be used as input to any melanoma classification algorithm. The paper explicitly notes that IHP can be 'easily combined with the majority of classification algorithms.' This plug-and-play compatibility means IHP can serve as a preprocessing stage for deep learning classifiers, providing a clean, artifact-reduced lesion ROI.
Genetic algorithm implementation: The optimization problem - finding the direction that maximizes bimodal histogram separation - is solved using a Genetic Algorithm (GA). GA works by maintaining a population of candidate projection vectors (chromosomes) that evolve through selection, crossover, and mutation over multiple generations. The population is restricted to directions that produce bimodal histograms, reducing the search space to biologically meaningful solutions.
GA parameters: Population size was set to 100 candidates with 200 evolutionary iterations. Standard GA operators were used: 90% of the best candidates were selected for reproduction, 70% underwent crossover, and 50% underwent mutation. Real-valued chromosome encoding was used rather than binary encoding, providing finer granularity in the optimization. This parameter set was validated to reliably find the optimal separation direction.
Computational cost: Processing one 576x768 RGB image required approximately 11 minutes on a Pentium IV 3 GHz processor (circa 2007). This computation time reflects the genetic algorithm search across all possible projection directions. For clinical deployment, this would need to be optimized through either more efficient optimization algorithms or modern GPU acceleration, which can provide 100-1000x speedups.
Comparison with alternatives: PCA - a faster linear transformation approach - was also evaluated but produced inferior segmentation because PCA maximizes variance rather than class separation. The principal component direction can shift unpredictably when artifacts like hair or blood vessels are present, while IHP's bimodal optimization directly targets the lesion-skin boundary regardless of other image structures.
Integration with classification systems: The paper explicitly positions IHP as a segmentation module to be integrated into complete CAD systems. Future work should validate IHP segmentation as a preprocessing step for deep learning classifiers, measuring whether more accurate segmentation translates to higher melanoma classification accuracy compared to other segmentation approaches.
Multi-lesion and total body photography: IHP was evaluated on single-lesion dermoscopy images. Modern total body photography (TBP) systems capture hundreds of skin lesions simultaneously. Extending IHP to automatically detect and segment all lesions in a total body image - identifying new, changed, or suspicious lesions across follow-up visits - represents a clinically important extension.
Speed optimization for real-time deployment: The genetic algorithm's iterative nature creates computational overhead. Alternative optimization strategies - gradient-based methods, simulated annealing, or direct neural network approximation of the IHP transform - could achieve real-time performance. Given the mathematical simplicity of the final linear transform, the computational burden is entirely in the optimization phase, which could be precomputed or approximated.
Extension to 3D dermoscopy and OCT: IHP operates on 2D images but the core concept - finding projections that maximize mode separation in pixel value histograms - extends naturally to 3D volumetric data from OCT or confocal microscopy. Applying IHP-like unsupervised band selection to 3D volumetric imaging could improve lesion volume estimation and depth characterization, which are important for melanoma staging.