Diagnostic Challenge Early and accurate classification of skin lesions is critical for improving melanoma survival rates. Manual dermoscopic analysis is time-consuming, requires specialized expertise, and suffers from inter-observer variability. Automated computer-aided diagnosis systems that can reliably segment and classify skin lesions are therefore a high-priority research goal.
End-to-End Pipeline This paper presents a fully automated pipeline for skin lesion analysis consisting of four sequential stages: image contrast enhancement, deep learning-based segmentation, hybrid deep feature extraction with optimization, and multi-class classification. Each stage was specifically designed to address known limitations in prior automated approaches.
Key Innovations The pipeline introduces several novel components: the LCcHIV contrast enhancement algorithm, a Deep Saliency Segmentation network, an Improved Moth Flame Optimization (IMFO) algorithm for feature selection, and a Multi-Model Classification Architecture (MMCA) fusion strategy combined with a Kernel Extreme Learning Machine (KELM) classifier.
Evaluation Benchmarks The system was evaluated on three established dermoscopy datasets: ISBI 2016 Challenge, PH2, and HAM10000. Performance was measured using standard metrics including accuracy, sensitivity, specificity, and Jaccard index to enable comparison with state-of-the-art methods.
Contrast Enhancement Motivation Dermoscopy images often contain low-contrast lesion boundaries, irregular illumination, and visual artifacts including hair, air bubbles, and gel reflections. Standard contrast enhancement methods may amplify noise while improving lesion visibility. The proposed LCcHIV (Local Contrast Combined with Histogram and Image Variance) algorithm addresses these limitations.
LCcHIV Algorithm Design The LCcHIV method combines local contrast enhancement with histogram equalization and image variance weighting to selectively enhance lesion-skin boundaries while suppressing artifact amplification. This adaptive approach applies different enhancement strengths based on local image statistics, preserving clinically relevant features.
Hair Artifact Removal A morphological hair removal preprocessing step using top-hat transform and inpainting was applied to dermoscopy images before contrast enhancement. This step reduces the segmentation-confounding effect of hair occlusion, which is particularly common in dermal lesion images and can cause segmentation algorithms to trace hair boundaries instead of lesion edges.
Preprocessing Impact Ablation experiments confirmed that LCcHIV preprocessing significantly improved downstream segmentation accuracy compared to using raw or standard histogram-equalized images. The enhanced images provided clearer lesion-background contrast that benefited the subsequent deep saliency segmentation network.
Network Architecture The Deep Saliency Segmentation (DSS) network is a 10-layer convolutional neural network designed to generate pixel-wise lesion segmentation masks. The architecture incorporates multi-scale feature extraction through parallel convolutional paths operating at different spatial resolutions to capture both fine boundary details and global lesion structure.
Saliency Map Generation The network produces saliency maps that highlight regions corresponding to skin lesions, leveraging spatial attention mechanisms to focus on lesion-relevant features. These saliency maps are thresholded and post-processed using morphological operations to generate clean binary segmentation masks.
Training Strategy The segmentation network was trained with augmented data including rotations, flips, and color jitter to improve robustness to orientation and imaging condition variability. The loss function combined binary cross-entropy with a Dice loss term to handle the class imbalance inherent in lesion segmentation, where lesion pixels are typically far fewer than background pixels.
Segmentation Performance The system achieved segmentation accuracy of 95.38% on ISBI 2016 and 98.70% on the PH2 dataset. These results exceeded multiple prior state-of-the-art methods. Performance on the more challenging HAM10000 dataset demonstrated robustness across diverse lesion types beyond the binary melanoma/non-melanoma classification.
Deep Feature Extraction Features were extracted from two pre-trained deep networks: ResNet101 and DenseNet201. ResNet101's residual connections enable extraction of deep semantic features while preserving gradient flow, while DenseNet201's dense connections produce a richer feature space. Features from both networks were extracted from the penultimate layer for maximum representational capacity.
Feature Fusion Strategy A Multi-Model Classification Architecture (MMCA) fusion strategy combined features from ResNet101 and DenseNet201 by concatenation, producing a high-dimensional feature vector. While this increases representational richness, it also introduces redundant and irrelevant features that can harm classifier performance, motivating the need for feature selection.
Improved Moth Flame Optimization The standard Moth Flame Optimization (MFO) algorithm simulates moths navigating by maintaining a fixed angle to moonlight. The IMFO modification introduced in this paper incorporates an adaptive position updating mechanism and Levy flight distribution to improve the search's exploration-exploitation balance, avoiding local optima common in high-dimensional feature selection problems.
Feature Selection Results IMFO selected a compact subset of the most discriminative features from the concatenated ResNet101-DenseNet201 feature space. Compared to using all features, IMFO-selected features improved classification accuracy while dramatically reducing computational complexity. The selected feature subset was more generalizable across the three evaluation datasets.
Kernel Extreme Learning Machine The Kernel Extreme Learning Machine (KELM) classifier was chosen for its computational efficiency and strong generalization properties. Unlike standard SVMs, KELM uses a kernel function to map input features to a high-dimensional space while maintaining single-pass learning, making it well-suited for the selected deep feature space.
Multi-Class Classification Results On the HAM10000 dataset with seven lesion categories, the system achieved 90.67% classification accuracy. This multi-class performance on a challenging real-world dataset with class imbalance demonstrates the pipeline's practical utility beyond binary melanoma detection.
Comparison with State-of-the-Art The proposed pipeline outperformed multiple prior methods on all three evaluation datasets. Specifically, it exceeded the performance of standalone ResNet101 and DenseNet201 architectures, demonstrating that the feature selection and fusion strategy adds measurable value over simply using a larger pre-trained network.
Ablation Study Findings Systematic ablation experiments confirmed the contribution of each pipeline component. Removing LCcHIV preprocessing reduced segmentation accuracy, removing IMFO feature selection reduced classification accuracy, and replacing KELM with a standard SVM reduced classification performance. Each innovation made an independent contribution to the system's overall performance.
Decision Support Integration The pipeline's fully automated nature from raw image to classification result makes it suitable for integration into clinical dermoscopy workflows as a decision support tool. Dermatologists could receive AI-generated segmentation overlays and classification probabilities alongside the dermoscopy image during examination.
Multi-Class Capability Unlike binary melanoma-vs-benign systems, the pipeline's ability to classify seven skin lesion categories from HAM10000 addresses the clinical reality that dermatologists face a spectrum of diagnoses. This multi-class capability better matches the clinical decision problem and reduces the risk of false confidence from simplified binary outputs.
Computational Efficiency The KELM classifier's single-pass learning and IMFO's reduced feature dimensionality contribute to a computationally efficient system. Reduced inference time is critical for clinical deployment in high-volume settings, and the pipeline's efficiency makes real-time analysis during clinical consultations feasible.
Teledermatology Potential The pipeline's automated end-to-end processing is well-suited for teledermatology applications where images are submitted remotely without a specialist present to guide image quality. The LCcHIV preprocessing's ability to handle variable image quality makes it particularly relevant for this use case.
Dataset Diversity The evaluation datasets, while standard benchmarks, represent a limited range of imaging conditions, patient demographics, and geographic populations. Future validation should incorporate datasets from diverse clinical settings, dermoscope types, and patient populations to assess real-world generalizability.
Clinical Validation Performance on benchmark datasets does not directly translate to clinical impact. Prospective clinical studies comparing AI-assisted versus unassisted dermatologist diagnostic accuracy, referral decisions, and patient outcomes are needed to establish the pipeline's actual clinical value.
Network Architecture Modernization The pipeline uses ResNet101 and DenseNet201, which were state-of-the-art at the time of publication. Incorporating newer architectures such as vision transformers or EfficientNet families, and exploring self-supervised pre-training on large dermatological image collections, could further improve feature quality.
Multimodal Extension Future work could extend the pipeline to incorporate clinical metadata including patient age, anatomical site, and lesion history alongside image features. Multimodal models that integrate visual and tabular data have shown improved performance in other medical AI applications and represent a natural evolution of this framework.