Clinical Context Skin cancer is one of the most common cancers globally, with melanoma responsible for the majority of skin cancer-related deaths. Early detection is critical - dermatologists with limited expertise show reduced diagnostic accuracy, and manual lesion analysis is time-consuming, subjective, and error-prone. Deep learning-based computer-aided diagnosis offers a consistent, rapid alternative.
Study Design This study proposes a deep learning pipeline for binary classification of skin lesions (benign versus malignant) using the ISIC 2018 dataset, comprising 3533 skin lesion images including benign, malignant, nonmelanocytic, and melanocytic tumor types. A novel preprocessing step using ESRGAN for image enhancement precedes both a custom CNN and multiple transfer learning models.
Key Innovation The primary novel contribution is the use of Enhanced Super-Resolution Generative Adversarial Network (ESRGAN) as a preprocessing step to improve image quality before classification. This approach, not previously applied to skin lesion classification, aims to enhance fine detail in dermoscopy images that may be diagnostically relevant but poorly preserved in standard image formats.
Models Compared Four classification models are compared: a custom-designed CNN, and three fine-tuned transfer learning models - ResNet50, InceptionV3, and Inception ResNet. This multi-model comparison on a standardized dataset enables direct performance benchmarking.
ESRGAN Introduction Enhanced Super-Resolution Generative Adversarial Network (ESRGAN) is a GAN-based image upscaling technique designed to generate perceptually realistic high-resolution images from lower-resolution inputs. Unlike traditional bicubic upscaling that produces blurry outputs, ESRGAN generates texturally rich, high-fidelity enhanced images that better preserve fine structural details.
Rationale for Dermoscopy Application Dermoscopy images vary in resolution and quality depending on acquisition hardware and imaging conditions. Compression artifacts and resolution limitations may obscure fine diagnostic features such as subtle color gradients, delicate pigmentation patterns, and lesion border texture. ESRGAN preprocessing aims to recover and enhance these potentially diagnostically relevant details.
Preprocessing Pipeline The preprocessing pipeline includes ESRGAN enhancement, followed by standard steps: data augmentation (rotations, flips, color jitter) to address the limited dataset size, normalization to standardize pixel value distributions across images, and resizing to the input dimensions required by each classification model.
Expected Impact The hypothesis is that ESRGAN-enhanced images provide richer feature representations for the convolutional networks to learn from, improving discrimination between fine-grained visual differences between benign and malignant lesions that may not be apparent in the original images.
Custom CNN Design A custom convolutional neural network was designed with multiple convolutional-pooling blocks for hierarchical feature extraction from dermoscopy images. The network was trained from scratch on the ISIC 2018 dataset, relying entirely on the dermoscopy domain images for feature learning rather than leveraging knowledge from natural image pre-training.
ResNet50 Transfer Learning ResNet50 is a 50-layer residual network pre-trained on ImageNet. Its residual connections enable stable training of deep architectures by providing identity shortcut paths that prevent gradient vanishing. Fine-tuning was performed by replacing the final classification layer with a two-class output and retraining all weights on the ISIC 2018 data with appropriate learning rate scheduling.
InceptionV3 Transfer Learning InceptionV3 uses multi-scale convolutional filters (1x1, 3x3, 5x5) in parallel inception modules, enabling simultaneous capture of features at different spatial scales. This multi-scale approach is well-suited to dermoscopy images where diagnostically relevant features span multiple spatial resolutions from fine texture to global lesion shape.
Inception ResNet Transfer Learning Inception ResNet combines the multi-scale feature extraction of inception modules with residual connections, providing both architectural advantages. The combined architecture offers a balance between feature richness and training stability, producing a strong baseline for dermoscopy classification tasks.
Custom CNN Performance The custom-designed CNN achieved 83.2% accuracy on the binary benign/malignant classification task. This competitive performance against established pre-trained architectures demonstrates that even a purpose-built smaller network can achieve reasonable classification accuracy with appropriate preprocessing and augmentation strategies.
ResNet50 Performance ResNet50 with fine-tuning achieved 83.7% accuracy, marginally outperforming the custom CNN. The residual architecture's deep feature learning capability provides a small but consistent advantage over a shallower custom design, consistent with findings from related skin lesion classification benchmarks.
InceptionV3 Best Performance InceptionV3 achieved the highest accuracy of 85.8% among all tested models. The multi-scale inception module architecture's ability to capture features at different spatial resolutions appears well-matched to the visual characteristics of dermoscopy images, where both fine texture and global shape are diagnostically relevant.
Inception ResNet Performance Inception ResNet achieved 84.0% accuracy, intermediate between InceptionV3 and ResNet50. While the combined architecture offers theoretical advantages from both inception and residual design principles, the performance gap over InceptionV3 alone suggests that residual connections provide diminishing returns on this dataset size.
Screening Tool Potential A binary benign/malignant classifier with 85%+ accuracy could function as a preliminary screening tool to prioritize dermoscopy images for expert review. In high-volume dermatology settings or teledermatology platforms where all images cannot receive immediate expert attention, AI-based triage could flag high-probability malignant lesions for urgent dermatologist review.
ESRGAN for Clinical Quality Improvement The ESRGAN preprocessing approach has applications beyond AI classification. Enhancing the visual quality of dermoscopy images before dermatologist review could improve the diagnostic quality of images captured with lower-quality devices, potentially relevant for teledermatology applications in resource-limited settings where high-end dermoscopes may not be available.
Dataset Scale and Clinical Relevance The ISIC 2018 dataset's 3533 images represent a relatively modest dataset by deep learning standards. Real-world clinical deployment would benefit from training on far larger and more diverse datasets to improve generalizability across different patient demographics, skin phototypes, and imaging conditions.
Benign/Malignant Binary Task The binary classification framing - benign versus malignant - is clinically relevant for the primary triage decision of whether to biopsy a lesion. However, clinical practice requires more nuanced differential diagnosis within malignant categories. Future extensions to multi-class classification distinguishing melanoma, basal cell carcinoma, squamous cell carcinoma, and pre-malignant lesions would increase clinical utility.
Dataset Limitations The ISIC 2018 dataset of 3533 images is smaller than many benchmark skin lesion datasets. Validation on larger, more diverse datasets including the ISIC 2019 and 2020 datasets, and international clinical collections, would better establish generalizability across populations and imaging conditions.
ESRGAN Contribution Quantification The study does not include an ablation experiment comparing model performance with and without ESRGAN preprocessing, making it difficult to quantify the preprocessing's independent contribution to classification performance. Future work should isolate the ESRGAN effect through controlled ablation to justify its computational cost.
Multi-Class Extension Extension from binary to multi-class classification covering the full ISIC 2018 disease spectrum (seven categories) would increase clinical relevance. Multi-class classification is more challenging due to class imbalance and overlapping visual features but provides more actionable differential diagnosis information.
Explainability Integration Deep learning classifiers function as black boxes, providing predictions without diagnostic reasoning. Integrating Grad-CAM or similar gradient visualization techniques to highlight dermoscopy image regions contributing to malignant predictions would improve interpretability for clinical users and support regulatory review for medical device approval.