Shallow Tuned DenseNet: A Lightweight Convolutional Neural Network Approach for Enhanced Skin Lesion Recognition

Skin lesions are irregularities on the skin, often harmless, though some may progress to skin cancer. Detecting them early is crucial for successful treatment and prevention of cancer development. Traditional detection methods involve significant expenses and testing. This study introduces an auto-mated system for classifying skin lesions using a Shallow Tuned DenseNet (STDN), a compact Convolutional Neural Network (CNN) model. We employ Transfer Learning (TL) by fine-tuning a pretrained DenseNet-169 model with additional lightweight layers, allowing the STDN to adapt to the PH² dataset's specific characteristics while remaining efficient. Initially, we utilize data augmentation techniques to address class imbalances within the dataset, improving model performance by reducing bias towards the majority class. The evaluation demonstrates a promising hundred percent accuracy with perfect F1-score and specificity. We attribute these results to effective transfer learning utilization, enhanced tuning, and partially to data augmentation, which mitigates overfitting and biases, facilitating robust feature extraction. Furthermore, we employ Grad-CAM (Gradient-weighted Class Activation Mapping) to visualize and interpret the model's predictions, enhancing trust and comprehension. The proposed STDN model makes a significant contribution by achieving high accuracy with a lightweight design, offering a potentially cost-effective and efficient solution for automated skin lesion detection, thereby aiding in early cancer prevention.