[Paper Review] A Machine Vision Approach to Preliminary Skin Lesion Assessments

Ali Khreis, Ro'Yah Radaideh+1 more|arXiv (Cornell University)|Jan 21, 2026

Cutaneous Melanoma Detection and Management0 citations

TL;DR

The paper integrates the ABCD dermoscopy rule with machine learning to assess skin lesions, comparing a handcrafted rule-based pipeline against ML models, including a CNN trained from scratch that improves accuracy on HAM10000 data.

ABSTRACT

Early detection of malignant skin lesions is critical for improving patient outcomes in aggressive, metastatic skin cancers. This study evaluates a comprehensive system for preliminary skin lesion assessment that combines the clinically established ABCD rule of dermoscopy (analyzing Asymmetry, Borders, Color, and Dermoscopic Structures) with machine learning classification. Using a 1,000-image subset of the HAM10000 dataset, the system implements an automated, rule-based pipeline to compute a Total Dermoscopy Score (TDS) for each lesion. This handcrafted approach is compared against various machine learning solutions, including traditional classifiers (Logistic Regression, Random Forest, and SVM) and deep learning models. While the rule-based system provides high clinical interpretability, results indicate a performance bottleneck when reducing complex morphology to five numerical features. Experimental findings show that transfer learning with EfficientNet-B0 failed significantly due to domain shift between natural and medical images. In contrast, a custom three-layer Convolutional Neural Network (CNN) trained from scratch achieved 78.5% accuracy and 86.5% recall on median-filtered images, representing a 19-point accuracy improvement over traditional methods. The results demonstrate that direct pixel-level learning captures diagnostic patterns beyond handcrafted features and that purpose-built lightweight architectures can outperform large pretrained models for small, domain-specific medical datasets.

Motivation & Objective

• Motivate early, interpretable lesion assessment to aid melanoma detection.
• Develop a hybrid pipeline that computes a Total Dermoscopy Score (TDS) from dermoscopic features.
• Compare handcrafted, classical ML, and deep learning approaches on a standardized dataset.
• Assess the impact of domain shift and dataset size on model performance.
• Demonstrate the viability of a lightweight, purpose-built CNN for small medical datasets.

Proposed method

• Implement a rule-based pipeline to compute a Total Dermoscopy Score (TDS) from ABCD rule features.
• Evaluate traditional classifiers (Logistic Regression, Random Forest, SVM) on the same task.
• Test transfer learning using EfficientNet-B0 on HAM10000-derived images.
• Train a custom three-layer CNN from scratch on median-filtered images.
• Compare performance in terms of accuracy and recall across approaches.

Experimental results

Research questions

• RQ1 Can a rule-based ABCD-derived TDS provide comparable diagnostic signals to ML models?
• RQ2 Do traditional ML models or transfer-learned networks outperform handcrafted features on HAM10000-derived data?
• RQ3 Can a small, purpose-built CNN trained from scratch outperform larger pretrained models on a small medical dataset?

Key findings

• The rule-based TDS approach offers high clinical interpretability but shows a performance bottleneck when compressing morphology into five features.
• Transfer learning with EfficientNet-B0 failed significantly due to domain shift between natural and medical images.
• A custom three-layer CNN trained from scratch achieved 78.5% accuracy and 86.5% recall on median-filtered images.
• The CNN surpassed traditional methods with about a 19-point accuracy improvement over them.
• Direct pixel-level learning captured diagnostic patterns beyond handcrafted features, supporting lightweight architectures for small datasets.