Research Paper: Optimized Deep CNN Architecture for Multi-dermoscopic Diseases Classification

 

Abstract

Skin cancer is one of the most common cancers all over the world. It is easily curable when it is detected in its beginning stage. Early detection of malignant through accurate techniques and innovative technologies has a great impact on decreasing mortality rates associated with this disease. However, an imbalance of recall measures between classes affected the performance of existing models.

This study proposes a method using deep convolutional neural networks aiming to classify skin lesion as a multi-class classification problem. It involves three major features, namely customized batch logic, customized loss function and optimized fully connected layers. The training dataset is kept up to date including 24,530 dermoscopic images of seven categories; this is the largest dataset by far. The performances of eight proposed combined methods are evaluated by a test dataset of 2,453 images. The best combination of EfficientNetB4-CLF achieved the highest accuracy at 88.83% and mean recall at 83.66%. Besides, the proposed deep learning architecture is suitable for multi skin disease classification, solved underfitting and avoided overfitting problems, and achieved improving performance when used with customized batch logic and loss function.

Keywords – Skin Lesion Classification, Melanoma, Deep CNN, Customized Loss, Balanced Batch Logic