The clinical problem Skin cancer affects one-third of all cancer cases globally, yet the initial appearance of different skin cancer subtypes can be nearly identical, making accurate differentiation challenging even for experienced dermatologists. Average diagnostic accuracy for dermoscopy ranges from 60-80%, rising to only 80% even with ten years of experience.
What SCDNet does This paper introduces SCDNet, a novel deep learning framework that classifies dermoscopy images into four major skin cancer types - Melanoma, Melanocytic Nevi, Basal Cell Carcinoma, and Benign Keratosis - achieving 96.91% overall accuracy on the ISIC 2019 benchmark dataset.
Why it matters Automated multi-classification systems that can simultaneously distinguish these four major categories could assist dermatologists in prioritizing worrying lesions, reducing unnecessary biopsies, and enabling earlier treatment - particularly important given that early-stage melanoma carries a greater than 96% survival rate compared to just 5% at advanced stages.
Dataset scale and composition SCDNet was trained on the ISIC 2019 dataset containing 25,331 dermoscopic images across four categories: 4,522 melanoma images, 12,875 melanocytic nevi, 3,323 basal cell carcinoma, and 2,624 benign keratosis images - a major benchmark in computational dermatology.
Data splits and normalization Images were divided 70:20:10 for training, testing, and validation respectively. All images were standardized to 224 x 224 pixel resolution and normalized using KerasImageDataGenerator to ensure consistent model training and prevent overfitting.
Class imbalance consideration The dataset contains a significant imbalance - melanocytic nevi outnumber all other classes combined. The study used this real-world distribution to train the model, reflecting the clinical reality where benign nevi vastly outnumber malignant lesions in practice.
Hybrid design philosophy SCDNet combines the pre-trained VGG16 backbone with two additional custom CNN blocks for feature extraction. VGG16 contributes learned visual representations from ImageNet pre-training, while the custom CNN layers specialize the model for skin lesion discrimination.
Convolutional layers and feature maps The initial convolutional layer applies 16 filters with a 3x3 kernel, with filter counts progressively increasing to 512 in deeper layers. Each convolution uses ReLU activation and zero-padding to preserve spatial dimensions, followed by max-pooling layers with 2x2 filters and stride 2 to reduce spatial dimensions and computational cost.
Classification head After the convolutional blocks, a flattening layer produces a one-dimensional feature vector. Two fully connected dense layers - the first with 512 neurons, the second with 4 neurons using SoftMax activation - produce class probabilities for the four skin cancer categories. Dropout regularization (value 0.20) prevents overfitting during training.
SCDNet accuracy leads all models Trained and tested over 50 epochs using stochastic gradient descent with initial learning rate 0.05, SCDNet achieved 96.91% accuracy, 92.18% recall, 92.19% precision, and 92.18% F1-score on the ISIC 2019 test set - outperforming all four comparison models.
Comparison models fall behind ResNet-50 achieved 95.50% accuracy, VGG-19 achieved 94.25%, AlexNet 93.10%, and Inception-V3 92.54% - a performance gap of 1.4 to 4.4 percentage points below SCDNet. The area under the ROC curve (AUC) for SCDNet was 0.9516, the highest among all tested architectures.
Class-specific performance SCDNet correctly classified 1,970 of 2,099 melanocytic nevi images, 550 of 622 basal cell carcinoma, 1,620 of 1,782 melanoma, and 530 of 563 benign keratosis images. Leave-one-out cross-validation (LOOCV) confirmed robust performance with 94.98% accuracy, validating generalizability.
Feature resolution advantage Standard pre-trained classifiers suffer from reduced spatial resolution in their final convolutional layers, causing loss of fine-grained discriminative features. SCDNet addresses this by incorporating merged dilated convolution values and appropriate 3x3 filter sizes that preserve feature resolution in the infected areas of dermoscopy images.
Noise and convergence advantages The SCDNet architecture speeds up training convergence while reducing the impact of structured noise in dermoscopy images - a critical issue because skin artifacts, hair, and imaging inconsistencies can confound standard classifiers. This leads to enhanced discriminative performance across all four cancer classes.
Clinical relevance Dermoscopy-based automated multi-classification directly addresses the dermatologist accuracy gap. Unlike binary classifiers (melanoma vs. benign) that dominate the literature, SCDNet handles the clinically realistic scenario where multiple cancer subtypes must be distinguished simultaneously, which is the actual task facing dermatologists.
Skin tone limitation The study's primary acknowledged limitation is that the ISIC 2019 dataset predominantly contains images of fair-skinned individuals. SCDNet has not been trained or validated on images of darker-skinned people, where skin cancer detection is already clinically more challenging - a significant equity gap that must be addressed.
Future dataset expansion The authors plan to extend SCDNet training to the ISIC 2020 dataset and other publicly available archives to improve generalizability across skin tones, imaging devices, and clinical settings.
Integration potential Deep learning models like SCDNet could be deployed as clinical decision support tools within dermatology workflows, providing a second opinion for suspicious lesions and reducing both unnecessary biopsies and missed malignancies. Further prospective clinical validation studies comparing model output against clinical outcomes are needed.