Hybrid Network Improves Skin Lesion Classification Accuracy

A research team published a hybrid feature-fusion network for dermoscopic skin lesion classification that reports strong cross-dataset performance, achieving 94.5% accuracy and 97% AUC-ROC. The pipeline pairs customized convolutional backbones with a Vision Transformer and a lightweight fusion classifier to capture both local texture and global context, and it adds Grad-CAM interpretability for clinician-facing explanations.

The system uses preprocessing steps including normalization and class-aware selective augmentation to address dataset imbalance and appearance variability. Feature extraction combines three customized encoders:

• •`VGG-16` with adaptive layer configuration to strengthen local texture filters
• •`ResNet-50` augmented with dermatological feature enhancement modules for contrast and border cues
• •`ViT` with dynamic patching to capture global attention patterns

Extracted feature vectors are concatenated and passed to `HDFNet`, a two-layer DNN classifier. Training and evaluation run on four public datasets: ISIC 2019, ISIC 2020, PAD-UFES, and DermQuest DERMIS. The workflow includes Grad-CAM visualizations to surface model attention and support clinical interpretability.

Medical imaging for skin cancer screening has long suffered from lesion heterogeneity, skin-type variation, and limited clinical interpretability. This paper embraces a hybrid strategy that explicitly combines convolutional texture priors with transformer-based global context, a pattern increasingly common in state-of-the-art vision work. Achieving 97% AUC-ROC across multiple datasets is notable because cross-dataset robustness is a harder, more clinically relevant target than single-dataset gains. The addition of class-aware augmentation and targeted dermatological feature enhancement helps bridge the domain gap between academic datasets and varied clinical images.

Validate reported metrics on external prospective cohorts and check whether the Grad-CAM explanations align with dermatologists' reasoning. Next steps should evaluate deployment constraints: inference latency, model calibration, and failure modes on darker skin types and rare lesion classes.