Enchancing Skin Cancer Classification Using EfficientNet-Based Architectures

Skin cancer ranks among the most prevalent cancers globally, presenting significant health challenges necessitating early and accurate detection for improved patient outcomes. Dermoscopy images play a crucial role in this domain, augmented by computer-based image analysis systems that enhance diagnostic accuracy. While Convolutional Neural Networks (CNNs) have shown promise in skin cancer classification, there remains a need for more effective models to enhance detection rates. This study addresses this gap by leveraging EfficientN et, a CNN variant known for its compound scaling technique, to improve skin lesion classification accuracy. We enriched the HAM10000 dataset using various augmentation techniques to identify optimal training strategies. Flip augmentation emerged as the most effective, achieving a classification accuracy of 80.78
^%
. Evaluating five EfficientN et models (B0 to B4) with flip augmentation revealed consistent minimum accuracies exceeding 79.04%. Fine-tuning by unfreezing convolutional base layers further enhanced performance, yielding a notable increase in minimum accuracy to 85.20%. Notably, EfficientNet B3 with flip augmentation and fine-tuning achieved the highest accuracy at 87.75
^%
, surpassing previous studies on the same dataset. This study underscores the effectiveness of our approach in advancing skin cancer classification accuracy, offering valuable insights and advancements in dermatological image analysis.