Blue-White Veil in Dermoscopy

• Blue-White Veil is a dermoscopic structure characterized by diffuse blue pigmentation with a semi-opaque white film, serving as a crucial indicator of melanoma.
• Classical computational approaches detect BWV using contextual pixel classification and aggregated shape descriptors to differentiate it from similar features.
• Recent deep learning methods with explainable AI components have improved BWV detection accuracy across diverse datasets, demonstrating robust clinical applicability.

Blue-white veil (BWV) is a dermoscopic structure defined as an “irregular, structureless area of confluent blue pigmentation with an overlying white ‘ground-glass’ film.” In dermoscopy, it is regarded as one of the most important features for the diagnosis of melanoma and has therefore become a target for automated image analysis, first through contextual pixel classification and lesion-shape analysis and, more recently, through deep convolutional models with explainable artificial intelligence (XAI) components (Celebi et al., 2010, Rasel et al., 10 Jul 2025).

1. Definition, appearance, and diagnostic role

BWV denotes a specific composite visual pattern rather than an arbitrary blue or white region. In dermoscopic practice, it appears as a diffuse, homogeneous blue-gray to blue-black zone overlaid by a semi-opaque whitish sheen. The blue component corresponds histologically to deeply situated melanin via the Tyndall effect, whereas the white component is due to orthokeratosis and dermal fibrosis (Celebi et al., 2010).

A central diagnostic issue is differentiation from other blue or white structures. Pure blue nevi are uniformly pigmented but lack the overlying “ground-glass” white layer and tend to be more regular and circumscribed. Regression structures, including white scar-like areas, are typically chalk-white and show pigment networks at their edge rather than homogeneous blue (Celebi et al., 2010). This distinction matters because a common misconception is to treat all blue-gray pigmentation as equivalent to BWV; the source literature instead frames BWV as a morphologically specific sign.

The diagnostic value of BWV is high in expert visual assessment. Among expert dermoscopists, it is reported as one of the strongest stigmata of invasive melanoma, with a sensitivity of 51% and specificity of 97% when assessed visually, as summarized in the computational study of Celebi et al. (Celebi et al., 2010). That study further treats BWV as a primary melanoma feature when it is the most characteristic clue to malignancy in a lesion.

2. Classical computational formulation in dermoscopy images

The 2010 arXiv report formalizes BWV detection as a machine-learning problem operating on dermoscopy images through preprocessing, contextual pixel classification, and lesion-level decision rules (Celebi et al., 2010). Preprocessing begins with lesion border masking: manually placed border points are fitted by a second-order B-spline and flood-filled to yield a binary lesion mask. Background skin color estimation then discards an outer “10%” band and uses the next “20%” band, defined by thresholding the Euclidean distance transform of the border, to compute mean skin color $(R_s,G_s,B_s)$. Non-skin pixels such as hairs, frames, and rulers are excluded using the rule $R>90 \wedge R>B \wedge R>G$.

Contextual pixel classification uses a $5\times5$ neighborhood around each pixel. For each feature $f$, the pixel’s value is taken to be the median over that window,

$$f_{\text{med}}=\operatorname{median}\bigl\{f(i,j)\mid (i,j)\in 5\times5\text{ window}\bigr\}.$$

A fast partial-sort, described as a “minimum exchange network,” computes the median in $\mathcal O(1)$ time per pixel (Celebi et al., 2010). This median-based contextualization is explicitly used to suppress noise and small artifacts.

The feature representation combines chromaticity, lesion-to-skin relative color, and texture. The absolute chromaticity coordinates are

$$F_1=\frac{R}{R+G+B},\quad F_2=\frac{G}{R+G+B},\quad F_3=\frac{B}{R+G+B}.$$

Relative color features compare lesion pixel values $(R_L,G_L,B_L)$ to skin values $(R_s,G_s,B_s)$ through ratios such as $F_4=R_L/R_s$, $R>90 \wedge R>B \wedge R>G$0, and $R>90 \wedge R>B \wedge R>G$1, as well as differences including $R>90 \wedge R>B \wedge R>G$2, $R>90 \wedge R>B \wedge R>G$3, and $R>90 \wedge R>B \wedge R>G$4; normalized sums and ratios extend these to twelve total relative color features $R>90 \wedge R>B \wedge R>G$5 (Celebi et al., 2010).

Texture is encoded using gray-level co-occurrence matrix (GLCM) descriptors averaged over $R>90 \wedge R>B \wedge R>G$6, $R>90 \wedge R>B \wedge R>G$7, $R>90 \wedge R>B \wedge R>G$8, and $R>90 \wedge R>B \wedge R>G$9 directions. The reported features are entropy,

$5\times5$0

contrast,

$5\times5$1

and correlation,

$5\times5$2

Classification is performed with a C4.5 decision tree. C4.5 recursively splits on one feature at a time using Information Gain Ratio. For a split on feature $5\times5$3 in dataset $5\times5$4,

$5\times5$5

with

$5\times5$6

To prevent overfitting, post-pruning is controlled by confidence factor $5\times5$7 and minimum examples per leaf $5\times5$8, with training performed on approximately 8,000 manually labeled veil versus non-veil pixels (Celebi et al., 2010).

3. Lesion-level inference from BWV extent and shape

The classical pipeline does not stop at pixelwise BWV labeling. It aggregates detected veil into lesion-level descriptors and then classifies lesions as melanoma or benign (Celebi et al., 2010). The first descriptor, $5\times5$9, is the fractional BWV area, defined as the area of detected BWV divided by the area of the lesion.

Two additional descriptors encode global lesion geometry. Circularity error, denoted $f$0, is derived from the lesion boundary and its centroid using the mean radius and the deviation of boundary-point radii from that mean. Ellipticity, denoted $f$1, is computed from second moments $f$2 via the eigenvalues $f$3 of the covariance matrix,

$f$4

and the resulting ellipticity score is defined from the ratio of the minimum to maximum eigenvalue, constrained to $f$5 (Celebi et al., 2010).

A second C4.5 tree takes $f$6 as input, although the final reported lesion-classification tree uses only $f$7 and $f$8. Its rule set is explicit: if $f$9, classify the lesion as benign; else if $$f_{\text{med}}=\operatorname{median}\bigl\{f(i,j)\mid (i,j)\in 5\times5\text{ window}\bigr\}.$$0, classify it as benign, a pattern described as common in blue nevi; otherwise classify it as melanoma (Celebi et al., 2010). This makes the method a hybrid of local appearance analysis and global-shape regularity constraints.

A plausible implication is that the shape stage functions as a specific corrective mechanism against overcalling uniformly bluish benign lesions. The source text makes this interpretation concrete by noting that benign blue-gray nevi with large uniform areas can still be misclassified if ellipticity $$f_{\text{med}}=\operatorname{median}\bigl\{f(i,j)\mid (i,j)\in 5\times5\text{ window}\bigr\}.$$1 fails to flag their regular shape (Celebi et al., 2010).

4. Validation, performance, and limitations of the classical pipeline

On held-out test pixels, the BWV detector reports sensitivity of 84.33% and specificity of 96.19%. At the lesion level, using the tree on $$f_{\text{med}}=\operatorname{median}\bigl\{f(i,j)\mid (i,j)\in 5\times5\text{ window}\bigr\}.$$2, the reported sensitivity is 69.35% and specificity is 89.97% over 545 dermoscopy images. In the subset of 133 melanomas in which BWV was a primary feature, sensitivity rises to 78.20% (Celebi et al., 2010). These are the principal validation figures for the decision-tree-based framework.

The same study defines sensitivity as

$$f_{\text{med}}=\operatorname{median}\bigl\{f(i,j)\mid (i,j)\in 5\times5\text{ window}\bigr\}.$$3

and specificity as

$$f_{\text{med}}=\operatorname{median}\bigl\{f(i,j)\mid (i,j)\in 5\times5\text{ window}\bigr\}.$$4

No formal confidence intervals or hypothesis tests are reported (Celebi et al., 2010). For an academic reader, this is an important methodological boundary: the reported operating characteristics are direct empirical outcomes on the study dataset rather than interval-estimated performance claims.

The limitations described in the source are closely tied to acquisition variability and lesion morphology. Chromaticity coordinates $$f_{\text{med}}=\operatorname{median}\bigl\{f(i,j)\mid (i,j)\in 5\times5\text{ window}\bigr\}.$$5 and relative color features are intended to compensate for differing illumination and skin tones, while median filtering in $$f_{\text{med}}=\operatorname{median}\bigl\{f(i,j)\mid (i,j)\in 5\times5\text{ window}\bigr\}.$$6 neighborhoods suppresses noise and small artifacts such as hairs and bubbles (Celebi et al., 2010). Nonetheless, very faint veils or strongly uneven lighting may yield false negatives. Benign blue-gray nevi with large uniform areas remain a failure mode when the regular-shape prior encoded by $$f_{\text{med}}=\operatorname{median}\bigl\{f(i,j)\mid (i,j)\in 5\times5\text{ window}\bigr\}.$$7 is insufficient.

The source also identifies directions for improvement: automate border detection to remove the manual step, incorporate multi-scale texture such as wavelets and learned features such as SVM or deep nets, and gather larger, more diverse image sets to support more robust classifier training and formal confidence interval estimation (Celebi et al., 2010).

5. Deep learning, annotation transfer, and explainable BWV detection

A later line of work replaces handcrafted contextual features and decision trees with a deep convolutional classifier designed specifically for BWV-positive versus non-BWV lesion categorization (Rasel et al., 10 Jul 2025). Because research into detecting BWV in dermatological images is described there as limited, the study first constructs annotations for non-annotated data using a color-threshold algorithm and then trains a DCNN with custom learnable activation layers.

The annotation procedure remains in RGB and adopts 80 human-perceived BWV color palettes from Madooei and Drew (2013). From these palettes, the per-channel extrema are reported as $$f_{\text{med}}=\operatorname{median}\bigl\{f(i,j)\mid (i,j)\in 5\times5\text{ window}\bigr\}.$$8, $$f_{\text{med}}=\operatorname{median}\bigl\{f(i,j)\mid (i,j)\in 5\times5\text{ window}\bigr\}.$$9, $\mathcal O(1)$0, $\mathcal O(1)$1, $\mathcal O(1)$2, and $\mathcal O(1)$3. BWV pixels therefore satisfy

$\mathcal O(1)$4

Each lesion image is tiled into non-overlapping $\mathcal O(1)$5 px patches; a patch is marked BWV if at least one pixel satisfies these inequalities, and the entire lesion is annotated as BWV-positive if at least one patch is flagged. A clinical dermatology expert adjudicated all annotations (Rasel et al., 10 Jul 2025).

The proposed network contains 31 layers, including repeated Conv→Norm→Pool blocks, a fully connected layer with 512 neurons, dropout with $\mathcal O(1)$6, a final fully connected layer with 2 neurons, and a softmax output over classes $\mathcal O(1)$7 (Rasel et al., 10 Jul 2025). Its custom activation is parametric ReLU (PReLU),

$\mathcal O(1)$8

where each $\mathcal O(1)$9 is trainable. This is contrasted in the source with Leaky-ReLU, which uses fixed $$F_1=\frac{R}{R+G+B},\quad F_2=\frac{G}{R+G+B},\quad F_3=\frac{B}{R+G+B}.$$0.

The study uses PH2, ISIC archive, and Derm7pt. PH2 contains 200 images with 120 BWV and 80 non-BWV cases, at $$F_1=\frac{R}{R+G+B},\quad F_2=\frac{G}{R+G+B},\quad F_3=\frac{B}{R+G+B}.$$1 px, with ground-truth BWV by expert. ISIC archive contributes 204 roughly balanced images, with 66 BWV and 138 non-BWV cases, spanning image sizes from $$F_1=\frac{R}{R+G+B},\quad F_2=\frac{G}{R+G+B},\quad F_3=\frac{B}{R+G+B}.$$2 to $$F_1=\frac{R}{R+G+B},\quad F_2=\frac{G}{R+G+B},\quad F_3=\frac{B}{R+G+B}.$$3 px; these are non-annotated images later annotated via the threshold algorithm plus clinical review. Derm7pt contains 2013 images, with 381 BWV and 1632 non-BWV cases, at $$F_1=\frac{R}{R+G+B},\quad F_2=\frac{G}{R+G+B},\quad F_3=\frac{B}{R+G+B}.$$4 px, with built-in annotations (Rasel et al., 10 Jul 2025).

Training uses stochastic gradient descent with momentum, momentum $$F_1=\frac{R}{R+G+B},\quad F_2=\frac{G}{R+G+B},\quad F_3=\frac{B}{R+G+B}.$$5, initial learning rate $$F_1=\frac{R}{R+G+B},\quad F_2=\frac{G}{R+G+B},\quad F_3=\frac{B}{R+G+B}.$$6, up to 250 epochs with early stopping on validation, batch size 32, shuffled training and validation data each epoch, and validation every 25 iterations. For binary classification, the loss is minibatch cross-entropy with $$F_1=\frac{R}{R+G+B},\quad F_2=\frac{G}{R+G+B},\quad F_3=\frac{B}{R+G+B}.$$7-weight decay:

$$F_1=\frac{R}{R+G+B},\quad F_2=\frac{G}{R+G+B},\quad F_3=\frac{B}{R+G+B}.$$8

6. Reported deep-learning performance, explainability, and current scope

The deep-learning study reports four cross-dataset experiments. Using the proposed DCNN with PReLU, the reported test-set results are: 85.71% accuracy, 71.43% precision, 83.33% sensitivity, 76.92% $$F_1=\frac{R}{R+G+B},\quad F_2=\frac{G}{R+G+B},\quad F_3=\frac{B}{R+G+B}.$$9, 86.67% specificity, and 81.75 AUC for AuP→ISIC (10%); 95.00% accuracy, 100.00% precision, 80.00% sensitivity, 88.89% $(R_L,G_L,B_L)$0, 100.00% specificity, and 89.60 AUC for AuIa→PH2 (10%); 95.05% accuracy, 73.68% precision, 100.00% sensitivity, 84.85% $(R_L,G_L,B_L)$1, 94.25% specificity, and 89.10 AUC for AuP+AuIa→Derm7pt (5%); and 90.00% accuracy, 60.00% precision, 100.00% sensitivity, 75.00% $(R_L,G_L,B_L)$2, 88.24% specificity, and 85.50 AUC for Derm7pt→AuP+AuIa (10%) (Rasel et al., 10 Jul 2025). Averaged across all four tests, the same model attains accuracy of 91.44%, precision of 76.28%, sensitivity of 90.83%, $(R_L,G_L,B_L)$3 of 81.42%, specificity of 92.29%, and AUC of 86.49%.

Explainability is provided through Local Interpretable Model-agnostic Explanations (LIME). For an input image $(R_L,G_L,B_L)$4, LIME approximates the black-box classifier $(R_L,G_L,B_L)$5 in the locality of $(R_L,G_L,B_L)$6 by a simple model $(R_L,G_L,B_L)$7:

$(R_L,G_L,B_L)$8

where

$(R_L,G_L,B_L)$9

and

$(R_s,G_s,B_s)$0

Images are segmented into a $(R_s,G_s,B_s)$1 superpixel grid, 5500 perturbed samples are generated, and a weighted linear model yields per-superpixel importance; bicubic interpolation then produces a smooth heatmap in which red regions represent the highest positive contribution to the BWV decision (Rasel et al., 10 Jul 2025).

The study states that the PReLU-DCNN outperforms Celebi et al. (2008), Madooei and Drew (2013), and Madooei et al. (2019) in every metric, with average gain of approximately 15–20% in accuracy and AUC, and that cross-dataset transfer demonstrates robust generalization (Rasel et al., 10 Jul 2025). At the same time, its scope is explicitly limited: the work focuses solely on BWV, other dermoscopic features such as streaks and regression remain to be integrated, and further clinical prospective trials are needed to validate performance in routine practice.

Taken together, these two arXiv works delineate the current technical framing of BWV in computational dermoscopy. The earlier framework models BWV as a contextual color-texture phenomenon whose lesion-level significance depends on extent and shape regularity (Celebi et al., 2010). The later framework shifts the task toward lesion categorization by presence of BWV, adds automated annotation transfer and XAI-based interpretability, and reports higher classification metrics across multiple datasets (Rasel et al., 10 Jul 2025). This suggests a broader methodological transition in BWV research from engineered feature spaces and explicit decision rules to dataset-centric deep models with post hoc explanation, while preserving the underlying clinical premise that BWV is a key dermoscopic sign of melanoma.