Automating the ABCD Rule for Melanoma Detection: A Survey

Abder-Rahman Ali, Jingpeng Li, Guang YangIEEE Access 2020 · OpenReview

Abstract

The ABCD rule is a simple framework that physicians, novice dermatologists and non-physicians can use to learn about the features of melanoma in its early curable stage, enhancing thereby the early detection of melanoma. Since the interpretation of the ABCD rule traits is subjective, different solutions have been proposed in literature to tackle such subjectivity and provide objective evaluations to the different traits. This paper reviews the main contributions in literature towards automating asymmetry, border irregularity, color variegation and diameter, where the different methods involved have been highlighted. This survey could serve as an essential reference for researchers interested in automating the ABCD rule.

Introduction

Melanoma is one of the most aggressive forms of skin cancer, yet it is highly curable when detected in its early stages. To aid in diagnosis, clinicians rely on the ABCD rule, which evaluates lesions based on Asymmetry, Border irregularity, Color variegation, and Diameter. However, manual application of this rule is inherently subjective, with diagnostic accuracy often ranging between 75% and 84% depending on the clinician's experience and fatigue.

Automating the ABCD rule through computer vision and machine learning aims to provide an objective, reproducible "second opinion." This survey explores the algorithmic evolution of these four pillars, transitioning from traditional geometric descriptors to modern deep learning frameworks.

The Four Pillars of Automated Detection

Asymmetry (A)

Asymmetry is a primary indicator of malignancy. Early automated methods employed "bi-fold" techniques, where a lesion image is rotated and reflected across its axes to calculate area differences. More sophisticated approaches now use Fourier descriptors to analyze boundary symmetry or Scale-invariant feature transform (SIFT) to compare texture similarity across lesion halves.

Illustration of lesion axes used for asymmetry and diameter calculation.

Border Irregularity (B)

Malignant lesions often exhibit jagged, ill-defined borders. Researchers quantify this "ruggedness" using fractal dimensions or the compactness index ($P^2/4\pi A$). Another technique, sigma-ratio analysis, uses scale-space filtering to identify structural indentations and protrusions that distinguish melanoma from smooth benign moles.

Comparison of skin lesion images and their extracted borders, highlighting differences in border irregularity.

Color Variegation (C)

Melanomas typically display a range of colors, including shades of brown, black, blue-gray, and white. Automation involves segmenting the lesion into color clusters using K-means or Kohonen maps. Modern methods often operate in the CIELab color space, which aligns more closely with human visual perception than standard RGB, allowing for more accurate distance calculations between pixels and suspicious color profiles.

Diameter (D)

Clinically, a diameter greater than 6mm is a warning sign. Algorithms calculate this by determining the Feret’s diameter (the distance between parallel tangents) or the major axis of a best-fit ellipse. Despite its simplicity, this metric is difficult to automate in public datasets that lack physical scale metadata (e.g., pixels-to-millimeter conversion factors).

Datasets and Performance

The development of these tools relies on high-quality public repositories such as the ISIC (International Skin Imaging Collaboration), which includes the massive HAM10000 dataset, and PH2, which provides 200 dermoscopic images with expert manual segmentations. Research shows that while individual traits like Asymmetry can achieve high accuracy (up to 93%), combining all four into a Total Dermoscopy Score (TDS) yields the best results, with reported sensitivities exceeding 91%.

Key Insights and Future Directions

A critical bottleneck in ABCD automation is segmentation—the process of isolating the lesion from the surrounding skin. Artifacts such as hair, air bubbles, and skin folds can significantly degrade performance. Furthermore, while Deep Learning (DL) has become the dominant paradigm, it often lacks the interpretability of "hand-crafted" features. Hybrid models that combine DL with traditional ABCD descriptors offer a promising path forward.

Future research must address the "Diameter" data gap by including physical scales in datasets and focus on making algorithms lightweight enough for mobile health applications, bringing expert-level screening to general practitioners and patients alike.

Summary updated

Abstract

The ABCD rule is a simple framework that physicians, novice dermatologists and non-physicians can use to learn about the features of melanoma in its early curable stage, enhancing thereby the early detection of melanoma. Since the interpretation of the ABCD rule traits is subjective, different solutions have been proposed in literature to tackle such subjectivity and provide objective evaluations to the different traits. This paper reviews the main contributions in literature towards automating asymmetry, border irregularity, color variegation and diameter, where the different methods involved have been highlighted. This survey could serve as an essential reference for researchers interested in automating the ABCD rule.
BibTeX
@inproceedings{ali2020automating,
  author = {Abder-Rahman Ali and Jingpeng Li and Guang Yang},
  title = {{Automating the ABCD Rule for Melanoma Detection: A Survey}},
  booktitle = {IEEE Access},
  year = {2020},
  url = {https://openreview.net/forum?id=muQKD5B9sP},
}
Figure