Source-linked AI summary
Lesion Border Detection in Dermoscopy Images
M. Emre Celebi, Hitoshi Iyatomi, Gerald Schaefer, William V. Stoecker
TL;DR
Automated lesion-border detection is an important but difficult component of dermoscopy analysis, motivating systematic evaluation of existing methods. This article reviews recent approaches and their computational and evaluation issues, concluding that standardized benchmarks and domain knowledge are important directions.
Problem
Automated lesion-border detection is important for dermoscopy analysis, but existing studies have problems with test-image acquisition, size, diagnostic distribution, evaluation, and method description.
Method
The article systematically overviews recent border detection methods, preprocessing, segmentation, postprocessing, computational issues, and evaluation practices.
Results
Recent studies used objective measures to assess validity, whereas earlier studies relied on visual assessment.
Takeaways & Limitations
A benchmark database should be prioritized to improve the quality of future automated border detection studies.
Takeaways & Limitations
Automated systems currently have limitations compared with experienced dermoscopy practitioners, and dermatologist border determination appears to use higher-level knowledge.
Abstract
from arXiv · showhide
Background: Dermoscopy is one of the major imaging modalities used in the diagnosis of melanoma and other pigmented skin lesions. Due to the difficulty and subjectivity of human interpretation, computerized analysis of dermoscopy images has become an important research area. One of the most important steps in dermoscopy image analysis is the automated detection of lesion borders. Methods: In this article, we present a systematic overview of the recent border detection methods in the literature paying particular attention to computational issues and evaluation aspects. Conclusion: Common problems with the existing approaches include the acquisition, size, and diagnostic distribution of the test image set, the evaluation of the results, and the inadequate description of the employed methods. Border determination by dermatologists appears to depend upon higher-level knowledge, therefore it is likely that the incorporation of domain knowledge in automated methods will enable them to perform better, especially in sets of images with a variety of diagnoses.
1 Introduction
Dermoscopy supports identification of clinically relevant lesion features, but automated border detection remains difficult and important for subsequent analysis. The article surveys recent methods and evaluation issues.
- Dermoscopy makes subsurface structures and morphological features more visible than conventional clinical images.
- Dermoscopy can improve differentiation between difficult lesions, although diagnostic accuracy may decrease for inexperienced dermatologists.
- Automated border detection is often the first step in dermoscopy image analysis and supports diagnosis and extraction of clinical features.Border-derived features include asymmetry, border irregularity, abrupt cutoff, pigment networks, globules, and blue-white areas.
- Border detection is challenging because lesions may have low contrast, irregular or fuzzy borders, artifacts, variegated coloring, and fragmentation.Reported artifacts include black frames, skin lines, blood vessels, hairs, air bubbles, and scar-like depigmentation.
- The article overviews recent border detection methods, including preprocessing, segmentation, postprocessing, and performance evaluation.
2 Preprocessing
Preprocessing prepares dermoscopy images for border detection through color transformation, contrast enhancement, and artifact removal. The choices involve calibration, representation, smoothing, computational cost, and artifact-specific methods.
- Preprocessing for border detection includes color space transformation, contrast enhancement, and artifact removal.
- Color processing may use scalar channels or multichannel representations, with color-space choices supporting channel reduction, luminance separation, or imaging invariance.
- CIEL*a*b* and CIEL*u*v* require careful acquisition-device calibration, which is frequently neglected in the literature.
- Independent histogram pursuit enhances contrast by transforming RGB images into a decorrelated space that maximally separates lesion and background skin.Border detection can then use simple clustering on the contrast-enhanced images.
- Artifact removal is necessary because frames, markings, bubbles, hairs, vessels, and skin lines reduce accuracy and increase computational time.General-purpose smoothing includes Gaussian, median, and anisotropic diffusion filters, while specialized methods target particular artifacts.
- Smoothing requires balancing mask size, edge preservation, filtering representation, and computational time.Excessively large masks blur edges; Gaussian and median filters can be implemented in constant time, whereas anisotropic diffusion depends on mask size and iterations.
3 Segmentation
Segmentation partitions dermoscopy images into homogeneous regions using several method families. Selection requires considering representation, automation, parameterization, and computational demands.
- Segmentation partitions an image into disjoint regions homogeneous with respect to a chosen property.
- Methods include histogram thresholding, clustering, edge-based, region-based, morphological, model-based, active-contour, and soft-computing approaches.
- Vector segmentation can require excessive computational time and makes choosing an appropriate color space difficult.
- Some methods are semi-automatic because active contours require an initial contour and seeded region growing requires initial region seeds.
- Increasing the number of segmentation parameters makes model selection harder.
4 Postprocessing
Postprocessing converts segmentation outputs into coherent lesion borders through merging, cleanup, smoothing, and expansion. These operations address fragmented regions, ragged contours, and systematic underestimation of dermatologist borders.
- Postprocessing is applied to segmentation outputs, with its precise sequence depending on the segmentation method.
- Region merging combines lesion subregions because segmentation often partitions heterogeneous lesions and background skin into multiple regions.
- Background color estimated from image corners can help eliminate subregions similar to background skin after black-frame removal.
- Small isolated regions can be removed with a binary area opening filter, while ragged borders can be smoothed using majority filtering, morphological filtering, or curve fitting.
- Computer-detected borders are often inside dermatologist-determined borders because automated methods detect sharper pigment changes than dermatologists.
- Border discrepancies can be reduced through morphological filtering, Euclidean distance transforms, or iterative region growing.
5 Evaluation
The article surveys subjective and objective evaluation of dermoscopy border detection, emphasizing disagreement among manual borders and limitations of common error measures. It also describes probabilistic and classifier-based alternatives for assessing detection quality.
- Evaluation approaches: Evaluation includes subjective dermatologist assessment and objective quantification of border errors against dermatologist-determined manual borders.Objective metrics commonly use true/false positive/negative pixel concepts and compare actual ground-truth pixels with detected output pixels.
- Manual-border variability: A single experienced dermatologist is not an absolute reference because manual borders show significant disagreement across dermatologists and repeated assessments.This motivates probabilistic rather than absolute gold-standard border models.
- Probabilistic evaluation: Guillod et al. modeled border uncertainty from fifteen border sets drawn by five dermatologists and computed automatic-border error using pixelwise misclassification probabilities.Each pixel received a misclassification probability based on how often it was selected as part of the lesion.
- Limitations of existing measures: Majority voting combines multiple manual borders but discounts their variation, whereas independent comparisons retain variation without producing a scalar error value.The latter makes comparisons among methods more difficult.
- Limitations of existing measures: An automated border entirely inside the manual borders can receive a very low error under Guillod et al.’s measure, showing that some metrics inadequately capture border variation.This illustrates why apparent agreement with one reference representation can mischaracterize detection quality.
- Probabilistic evaluation: The Normalized Probabilistic Rand Index accounts for manual-border variation and found four evaluated methods differed less than the commonly used XOR measure indicated.Its formulation is described as involved and referred to elsewhere in the article.
- Downstream evaluation: Border metrics do not quantify how detection errors affect classifier accuracy, which can instead be measured by comparing classification accuracy with manual versus automatic borders.Visual assessment also lacks the objective quality needed for parameter tuning and method comparisons.
6 Comparisons and Discussion
The reviewed border-detection studies vary in operating mode, image representation, preprocessing, segmentation, and evaluation. The literature also shows recurring weaknesses in test-set design, reporting, comparison, and the incorporation of clinically relevant knowledge.
- Method characteristics: 11/18 methods operate on multiple color channels, and 12/18 use a smoothing filter.These characteristics are reported among the reviewed border-detection methods.
- Method characteristics: Clustering is the most popular segmentation method, probably because robust algorithms are available.
- Evaluation and reporting: 9/18 studies rely on borders determined by a single dermatologist, limiting evaluation against variation among manual borders.The review recommends evaluating results using borders determined by multiple dermatologists.
- Evaluation and reporting: Only 5/18 studies report diagnostic distributions, and 8/18 do not compare results with another automated method.Diagnostic classes differ in border-detection difficulty, while comparison is hindered by unavailable public software and the lack of a public dermoscopy image database.
- Evaluation and reporting: Recent studies increasingly use objective measures, but XOR remains the most popular error function despite difficulties extending it to multiple manual borders.The review also calls for sufficiently detailed descriptions of image acquisition, border-detection procedures, and computational requirements.
- Discussion and implications: A public benchmark database is prioritized, and higher-level knowledge is identified as likely necessary for improved automated borders across varied diagnoses.Dermatologists incorporate diagnosis-dependent features such as halos, pale rims, reactive erythema, collision tumors, and scar-like regression when defining borders.