Skin Lesion Classification using EfficientNet B0 and B1 via Transfer Learning for Computer Aided Diagnosis

Automatic diagnosis of skin cancer from skin lesions using dermoscopy images has still been a challenging task for artificial intelligence, especially in Artificial Neural Networks using deep learning. Currently, previous research on skin lesion dermoscopy classification models have shown competitive performances whose accuracy remains to be improved. Applying and configuring the correct deep learning architecture has been an important factor in making an accurate automatic diagnosis. Hence, in this research, a new classification model was developed by using the latest architectures i.e., EfficientNet-B0 and B1 via transfer learning approach. This research was conducted by using HAM10000 dataset. To avoid overfitting due to the imbalanced data, the data were trained under some scenarios varying five variations of augmentation, class weight, and normalization within each model. The result showed that for EfficientNet-B0, the best model was the one that used only augmentation, which has the accuracy, precision, recall, and F1-scores of 91%, 76%, 68%, and 71%, respectively. While for EfficientNet-B1, the best model used augmentation and class weights with accuracy, precision, recall, and F1-score of 89%, 78%, 73%, and 73%, respectively. The best EfficientNet-B1 model can outperform the existing state-of-the-art model with an increase in recall and F1-score by 2% and 12% from the semi-supervised model, respectively. The model can also be integrated with a graphical user interface for dermatologists to use in dermoscopy examinations.