The Scale of the Problem Skin cancer is the most common cancer in humans, with melanoma accounting for only about 1% of cases but causing a disproportionately high number of deaths. Melanoma develops when melanocytes begin growing uncontrollably and, if not caught early, can spread to other organs and become fatal.
Limitations of Traditional Diagnosis Doctors have historically relied on biopsy - physically removing tissue from a suspicious lesion - to confirm whether it is cancerous. This process is painful, time-consuming, and can delay treatment. Non-invasive computer-based techniques that analyze skin lesion images offer a faster and more comfortable alternative.
What Deep Learning Brings Deep learning is a branch of artificial intelligence inspired by how the human brain processes information. In recent years, deep learning systems have achieved impressive results in image recognition tasks, making them well-suited to distinguish cancerous from benign skin lesions based on visual features like symmetry, color, size, and border irregularity.
Scope of This Review This paper systematically reviewed 51 research studies published between 2011 and 2021, analyzing four major deep learning approaches: artificial neural networks (ANN), convolutional neural networks (CNN), Kohonen self-organizing neural networks (KNN), and generative adversarial networks (GAN).
Systematic Search Process The authors searched five major databases - IEEE Xplore, Google Scholar, ACM DL, Springer, and Science Direct - using structured keyword combinations related to skin cancer, deep learning, and neural networks. The initial search returned 1,483 papers.
Multi-Stage Filtering Papers were narrowed down through title review (95 selected), abstract review (64 selected), and full-text evaluation (51 final papers). Quality assessment asked whether each study adequately covered deep learning techniques, was published in a reputable venue, and answered the core research questions.
Quality Scores The quality checks produced solid results: 77% of papers met topic coverage criteria, 82% met publication quality standards, and 79% adequately answered the review's research questions - indicating a high-quality, relevant body of literature.
Research Questions Addressed The review focused on two central questions: what are the major deep learning techniques used for skin cancer detection, and what are the main characteristics of datasets available for skin cancer research?
How ANNs Work Artificial neural networks mimic the brain's structure using layers of interconnected nodes. An input layer receives raw image data, hidden layers extract progressively more complex features, and an output layer assigns a class label - for example, 'melanoma' or 'benign.' The network learns by adjusting connection weights through a process called backpropagation.
ABCD Rule Integration One notable ANN-based system used the clinical ABCD rule - Asymmetry, Border irregularity, Color variation, and Diameter - to extract features before classification. This system, trained on 200 dermoscopic images, achieved 97.51% accuracy in distinguishing common moles, uncommon moles, and melanomas.
Ensemble Models for Better Accuracy Xie and colleagues combined a backpropagation neural network with a fuzzy neural network in an ensemble model, achieving 94.17% accuracy with sensitivity of 95% and specificity of 93.75%. Ensembles work by pooling predictions from multiple models, reducing the chance of misclassification.
Learning Algorithm Comparisons Masood and colleagues tested three ANN learning algorithms on the same dataset. The Levenberg-Marquardt algorithm achieved the highest specificity (95.1%) while the Scaled Conjugate Gradient method showed better sensitivity (92.6%) when trained for more epochs, illustrating how algorithm choice affects the clinical tradeoff between catching cancers and avoiding false alarms.
Why CNNs Excel at Images Convolutional neural networks are specifically designed for image analysis. Rather than treating each pixel independently, CNNs apply filters across the image to detect local patterns - edges, textures, color gradients - and build up a hierarchical understanding of the image. This makes them far more efficient and accurate than general ANNs for skin lesion images.
Transfer Learning Boosts Performance Many high-performing systems start with a CNN pre-trained on millions of general images (like ImageNet) and then fine-tune it on skin lesion data. Architectures such as ResNet-152, VGG-16, AlexNet, and Inception-v3 have all been adapted for melanoma classification, with the best systems achieving AUC values above 0.98.
Expert-Level Benchmarks One landmark CNN study achieved 72.1% classification accuracy against 21 certified dermatologists, with the deep learning model performing at or above expert level. Another study using DenseNet 201 achieved an ROC AUC of 98.79% across multiple skin lesion categories, outperforming highly trained dermatologists by at least 11% in mean results.
Segmentation Supports Classification Accurate identification of the lesion boundary - separating the lesion from surrounding skin - improves downstream classification. Systems that combined a fully convolutional residual network (FCRN) for segmentation with SVM or softmax classifiers achieved accuracy up to 94.9%, demonstrating that better segmentation leads to better diagnosis.
Diverse CNN Architectures Approaches range from lightweight models optimized for mobile phones (LightNet, 81.6% accuracy) to hybrid systems combining CNNs with autoencoders, recurrent networks, or fuzzy C-means clustering (achieving up to 98% accuracy with 93% Jaccard index), showing the versatility of the CNN framework.
Kohonen Self-Organizing Networks The Kohonen self-organizing map (KNN) is an unsupervised neural network that clusters data without needing labeled examples. It maps high-dimensional data - like the many features extracted from a skin image - into a lower-dimensional representation, preserving the relative distances between data points.
KNN Performance A KNN-based system using fluorescence spectroscopy data classified melanoma, nevus, and normal skin with only 2-3% error, outperforming traditional ANN (3-4% error). Another system combining a self-organizing map with a radial basis function (RBF) network achieved 93.15% accuracy across three cancer types, far exceeding naive Bayes (56.16%) and standard k-nearest neighbor classifiers (71.23%).
Generative Adversarial Networks for Data Augmentation GANs consist of two competing networks - a generator that creates synthetic images and a discriminator that tries to tell real from fake. In skin cancer research, GANs address a major problem: the shortage of labeled dermoscopic images needed to train deep learning models. By generating realistic-looking synthetic lesion images, GANs expand the effective training dataset.
GAN Classification Results A GAN-augmented system from Rashid and colleagues achieved 86.1% accuracy on seven skin lesion categories, outperforming ResNet-50 (79.2%) and DenseNet (81.5%). A progressive GAN (PGAN) method improved accuracy from 67.3% without augmentation to 70.1% with it, confirming that synthetic data can meaningfully boost classifier performance even when real data is limited.
ISIC Archive The International Skin Imaging Collaboration (ISIC) archive is the most widely used resource, containing over 25,000 dermoscopic images across eight lesion categories including melanoma, basal cell carcinoma, and seborrheic keratosis. Released annually with growing size and diversity, ISIC has become the benchmark dataset for comparing AI systems.
HAM10000 The Human Against Machine dataset with 10,000 training images (HAM10000) was collected over 20 years from two clinical centers in Australia and Austria. It contains 10,015 dermoscopic images spanning seven lesion types and specifically addresses the lack of diversity in earlier datasets, including images digitized from photographic prints dating back to before digital cameras were common.
PH2 Dataset The PH2 dataset contains 200 dermoscopic images from a Portuguese hospital, divided into 80 common nevi, 80 atypical nevi, and 40 melanomas. It includes clinical annotations such as pathologist-drawn lesion segmentations and assessments of dermoscopic criteria like pigment network, streaks, and regression areas - making it valuable for detailed feature analysis.
DermQuest, DermIS, and AtlasDerm Additional public datasets each offer different strengths: DermQuest provided 22,082 tagged images before its 2019 deactivation; DermIS contains 6,588 images with differential diagnoses and case reports; and AtlasDerm provides 1,024 high-quality images paired with educational content originally designed to help physicians recognize dermoscopic criteria for melanoma.
Dataset Bias and Skin Tone A critical limitation is that standard dermoscopic datasets predominantly contain images of light-skinned patients from Europe, Australia, and the United States. Neural networks trained on such data may not perform equally well on darker-skinned patients, creating potential healthcare disparities unless diverse training data is developed.
Class Imbalance and Small Lesion Sizes Real-world datasets are highly imbalanced - with hundreds of images for common cancer types but only a few for rarer ones. Small lesions (1-2mm) are also substantially harder to classify correctly, as the diagnostic accuracy for early-stage small lesions drops well below the 95-96% seen for larger lesions.
Computational Requirements and Age Gaps Training deep neural networks requires substantial GPU computing power, limiting access for smaller healthcare institutions. Standard datasets also skew toward younger patients, while several skin cancers (like BCC and Merkel cell carcinoma) predominantly affect people over 65, creating a gap in model generalizability for older populations.
Future Directions The review highlights several promising avenues: full-body photography systems to automate lesion discovery rather than classifying single submitted images; auto-organization techniques that learn image structure without labels; and integration of genetic and environmental risk factors (like UV exposure history or family history of melanoma) with image-based models for more holistic risk assessment.