Simulation and analysis of optical skin lesion images

Area of Science:

• Medical image analysis
• Computational dermatology
• Image processing algorithms

Background:

• Accurate analysis of skin lesion images requires ground truth data, which is often imprecise in clinical settings.
• Precise knowledge of lesion boundaries, skin patterns, and colors is crucial for evaluating image analysis algorithms.
• Existing clinical images lack the necessary ground truth for rigorous algorithm assessment.

Purpose of the Study:

• To develop a skin/lesion image simulation method that provides known ground truth characteristics.
• To create a synthetic dataset for validating feature estimation algorithms used in lesion classification.
• To enable accurate assessment of algorithms for differentiating malignant from benign skin lesions.

Main Methods:

• Synthesized monochrome and color skin/lesion images with known boundary, color, and skin pattern.
• Modeled skin and lesion textures using an auto-regressive (AR) process.
• Incorporated image artifacts like hair and specular reflections, and simulated inflammation areas.

Main Results:

• Developed image pre-processing techniques including AR model interpolation for hair removal and multiple illumination processing for specular reflection reduction.
• Extended a fast snake algorithm for accurate detection of skin lesion and inflammation boundaries.
• Identified skin line direction as a key feature to quantify skin pattern disruption by lesions.

Conclusions:

• Investigated the simulation of monochrome and color skin/lesion images as a viable alternative to real clinical data.
• Demonstrated the utility of synthetic images for validating image pre-processing, segmentation, and skin pattern analysis algorithms.
• Provided a method to generate image sets with known characteristics for algorithm validation in dermatology.