Exploring Augmentation Strategies for Balanced Skin Lesion Classification: An Explainable Lightly Tuned DenseNet 169 Architecture

Skin lesions indicate abnormalities in the skin, while skin cancer results from uncontrolled cellular proliferation due to cellular injury. Detecting these anomalies timely is crucial to prevent cancer progression. However, conventional methods are often resource-intensive and time-consuming, requiring sophisticated equipment. In response, this paper introduces an automated system that utilizes a Convolutional Neural Network (CNN) integrated with Transfer Learning (TL). The research focuses on fine-tuning the DenseNet 169 model to improve learning efficacy, achieving an empirical accuracy of ${9 8. 4 2 \%}$ on the ISIC 2020 dataset. Additionally, our approach implements various augmentation strategies to address dataset imbalances, with Prior Augmentation (PA) reducing bias toward the majority class, as demonstrated by the confusion matrix. While alternative augmentation methods yield similar outcomes, they tend to favor the majority class. Furthermore, this study involves identifying the responsible regions for lesion detection using the Gradient Class Activation Map (GradCAM) visualization method. This visualization enhances the explainability of the black-box model by pinpointing the regions crucial for decision-making.