Source-linked AI summary
Automatic Detection of Blue-White Veil and Related Structures in Dermoscopy Images
M. Emre Celebi, Hitoshi Iyatomi, William V. Stoecker, Randy H. Moss, Harold S. Rabinovitz, Giuseppe Argenziano, H. Peter Soyer
TL;DR
Melanoma-relevant blue-white veil is difficult to assess without dermoscopy, and systematic automated detection remained limited. The paper applies contextual pixel classification with a decision tree to detect veil and related structures, achieving 69.35% sensitivity and 89.97% specificity on 545 images.
Problem
Systematic detection of the melanoma-associated blue-white veil in dermoscopy images had not been published, despite dermoscopy revealing features invisible to naked-eye examination.
Method
The method combines preprocessing, feature extraction, contextual pixel classification, decision-tree induction, and lesion classification based on detected blue-white areas and shape.
Results
69.35% sensitivity and 89.97% specificity were obtained on 545 dermoscopy images.
Takeaways & Limitations
The presented blue-white veil detector processes a 768 x 512 image in a fraction of a second on an Intel Pentium D 2.66GHz computer.
Takeaways & Limitations
The study’s detection problem was constrained by reliance on manually determined lesion borders for separating feature extraction from automated border detection.
Abstract
from arXiv · showhide
Dermoscopy is a non-invasive skin imaging technique, which permits visualization of features of pigmented melanocytic neoplasms that are not discernable by examination with the naked eye. One of the most important features for the diagnosis of melanoma in dermoscopy images is the blue-white veil (irregular, structureless areas of confluent blue pigmentation with an overlying white "ground-glass" film). In this article, we present a machine learning approach to the detection of blue-white veil and related structures in dermoscopy images. The method involves contextual pixel classification using a decision tree classifier. The percentage of blue-white areas detected in a lesion combined with a simple shape descriptor yielded a sensitivity of 69.35% and a specificity of 89.97% on a set of 545 dermoscopy images. The sensitivity rises to 78.20% for detection of blue veil in those cases where it is a primary feature for melanoma recognition.
1. INTRODUCTION
Dermoscopy reveals melanoma-relevant structures invisible to the naked eye, motivating automated detection of blue-white veil, an important dermoscopic indicator whose systematic detection had not been studied.
- Clinical motivation: Dermoscopy visualizes features of pigmented melanocytic neoplasms that cannot be discerned by naked-eye examination.The modality offers higher diagnostic accuracy than observation without magnification when practiced by experienced observers.
- Clinical motivation: Blue-white veil comprises irregular, structureless confluent blue pigmentation covered by a white ground-glass film.It is identified as one of the morphological features available through dermoscopy.
- Clinical motivation: Blue-white veil is a significant dermoscopic indicator of invasive malignant melanoma, with 51% sensitivity and 97% specificity.These values describe the feature’s reported diagnostic performance.
- Research gap: Dermoscopy feature extraction remained relatively unexplored, with prior studies addressing pigment networks, globules, dots, and blotches.The cited literature included pilot and systematic studies of several features.
- Contribution: The article presents a machine-learning approach for detecting blue-white veil in dermoscopy images.The paper’s stated approach includes preprocessing, feature extraction, pixel classification, and lesion classification.
INSERT FIGURE 2 HERE
The study used 545 atlas-derived dermoscopy images with manually supervised lesion processing and balanced pixel sampling, while dermatologists annotated veil-related lesion characteristics.
- Image set: The image set contained 545 digital dermoscopy images obtained from two atlases.Images came from the Interactive Atlas of Dermoscopy and an American Academy of Dermatology DVD.
- Image set: Lesions were diagnosed histopathologically when significant melanoma risk was present and otherwise by follow-up examination.This describes the diagnostic basis used for the image set.
- Preprocessing: Lesion borders were manually determined under experienced-dermatologist supervision to separate feature extraction from automated border detection.Borders were constructed from selected points connected by a second-order B-spline and filled to form a binary mask.
- Preprocessing: Background skin color was estimated from a region outside the lesion after excluding a peripheral area to reduce inflammation and border-determination effects.Non-skin pixels and selected peripheral regions were excluded before averaging background color.
- Pixel sampling: Training and test pixels were randomly sampled from manually selected veil and non-veil regions to balance the two classes.The regions were selected across 100 images, including images with sizeable pure veil or non-veil areas.
- Clinical annotation: Dermatologists labeled primary blue-white veil and related structures, including gray or blue-gray veils and veils lacking the classic whitish film.A primary feature was the lesion feature considered most characteristic, recognizable, and specific for melanoma.
3. FEATURE EXTRACTION
The method used contextual pixel classification with neighborhood-based color and texture features designed to reduce illumination, rotation, and individual skin-color variation.
- Contextual features: Contextual pixel classification incorporated each pixel’s spatial neighborhood rather than treating pixels in isolation.Features were extracted in a 5 x 5 neighborhood and summarized by the neighborhood median.
- Feature composition: Fifteen color features and three texture features characterized the image pixels.Texture descriptors were based on gray-level co-occurrence matrices.
- Color features: Chromaticity coordinates F1, F2, and F3 were used because they are invariant to illumination direction and intensity.This invariance was considered essential for images acquired under uncontrolled conditions.
- Color features: Relative color compared a lesion pixel with the average background-skin color and provided 12 extracted features.These features were intended to compensate for illumination, digitization, and inter-individual skin-color variation.
- Texture features: Texture characterization used entropy, contrast, and correlation from normalized GLCMs across four directions.The selected statistics were gray-level shift-invariant, and directional values were averaged for rotation invariance.
4. PIXEL CLASSIFICATION
A pruned C4.5 decision tree classified pixels as veil or non-veil using interpretable rules, achieving strong manually selected-pixel performance while retaining only two features.
- Classifier: The C4.5 decision tree classified image pixels into veil and non-veil classes.Decision trees were chosen for interpretable rules and fast training and application.
- Classifier: C4.5 pruning was used to reduce overfitting and improve generalization to unseen data.Pruning replaces subtrees contributing little to predictive accuracy with leaves.
- Pixel-level results: 84.33% sensitivity and 96.19% specificity were obtained on manually selected test pixels.Sensitivity measured correctly detected veil pixels, while specificity measured correctly detected non-veil pixels.
INSERT FIGURE 5 HERE
The method applies induced decision rules to dermoscopy images, creates and smooths binary veil masks, and uses detected veil extent with lesion shape features for classification.
- Postprocessing: A 5 x 5 majority filter smoothed the initial veil masks by assigning each pixel its neighborhood’s majority class.The filter was applied to smooth mask borders.
- Detection results: The method detected most blue-white areas accurately in sample melanomas, a Reed/Spitz nevus, and a blue nevus.The sample results include melanomas in panels (a)–(f), a Reed/Spitz nevus in (g), and a blue nevus in (h).
- Lesion characterization: The detected blue-white veil area was normalized by lesion area to form the S1 feature.S1 is the area of detected blue-white veil divided by the area of the lesion.
- Lesion classification: S1 alone could misclassify blue nevi as melanoma, so circularity and ellipticity features were added to characterize lesion shape.The rationale was that benign lesions with blue-white areas may be highly circular or elliptical.
- Lesion classification: A C4.5 decision tree with 10-fold cross-validation classified lesions using S1, S2, and S3.S2 and S3 represent shape descriptors used alongside the blue-white area feature.
INSERT FIGURE 8 HERE
The lesion classifier combines blue-white veil extent with shape rules to distinguish melanoma from benign lesions, achieving different performance across the full set and targeted melanoma subset.
- Decision rules: Lesions with none to less than 0.9% blue-white areas were classified as benign.The decision tree checks ellipticity when lesions contain significantly larger blue-white areas.
- Decision rules: Lesions with large blue-white areas were classified as benign when S3 exceeded 0.979, otherwise as melanoma.S3 is the ellipticity feature used by the induced decision tree.
- Feature selection: The circularity feature S2 was discarded by induction, possibly because the more general ellipticity feature S3 captured its characteristics.This is presented as a possible explanation for S2’s removal.
- Performance: 69.35% sensitivity and 89.97% specificity were obtained on the entire set of 545 images.Sensitivity denotes correctly classified melanomas and specificity denotes correctly classified benigns.
- Performance: 82.94% overall classification accuracy was obtained for all areas, including blue-white-veil-related structures.The included related structures included blue-gray or gray veil.
- Subset performance: 76.87% sensitivity and 75.00% specificity were obtained on 44 benign and 134 melanoma cases with blue-white veil areas.This result is reported for the subset containing blue-white veil areas.
- Subset performance: 78.20% sensitivity was obtained for 133 melanomas in which blue-white veil was the primary feature.This is a sensitivity result for the primary-feature melanoma subset.
6. CONCLUSIONS
The study presents a machine-learning blue-white veil detector using preprocessing, feature extraction, decision-tree induction, rule application, and postprocessing.
- Contribution: The study describes a machine-learning approach for detecting blue-white veil in dermoscopy images.The method targets blue-white veil and related structures.
- Method: The method comprises preprocessing, feature extraction, decision-tree induction, rule application, and postprocessing.These stages form the detector’s processing pipeline.
- Method: Detected blue-white areas were characterized numerically and combined with ellipticity to classify 545 dermoscopy images.The combined features produced the reported lesion-level classification results.
- Results: 69.35% sensitivity and 89.97% specificity were achieved on 545 dermoscopy images.These are the reported lesion-classification performance values.
- Runtime: The detector processed a 768 x 512 image in a fraction of a second on an Intel Pentium D 2.66GHz computer.The reported runtime refers to the blue-white veil detector.
FIGURE LEGEND
The figures trace the blue-white veil detection workflow from preprocessing and pixel classification through postprocessing and image-level classification.
- Detection example: Figure 1 demonstrates the blue-white veil detection procedure on a melanoma dermoscopy image.The figure pairs a clinical image with a dermoscopy image.
- Workflow: Figure 2 presents an overview of the detection approach.
- Preprocessing: Figure 4 illustrates preprocessing, including border approximation, border masking, distance-based regions, and manually selected veil and non-veil regions.
- Pixel classification: Figure 5 shows the pixel classification tree used in the detection procedure.
- Postprocessing: Figure 6 compares initial and final veil masks after postprocessing.
- Classification results: Figures 7 and 8 present sample detection results and the image classification tree, respectively.Figure 7 distinguishes veil and non-veil borders with thick and thin lines.