AI outperformed every dermatologist: Improved dermoscopic melanoma diagnosis through customizing batch logic and loss function in an optimized Deep CNN architecture
Melanoma, one of most dangerous types of skin cancer, re-sults in a very high mortality rate. Early detection and resection are two key points for a successful cure. Recent research has used artificial intelligence to classify melanoma and nevus and to compare the assessment of these algorithms to that of dermatologists. However, an imbalance of sensitivity and specificity measures affected the performance of existing models. This study proposes a method using deep convolutional neural networks aiming to detect melanoma as a binary classification problem. It involves 3 key features, namely customized batch logic, customized loss function and reformed fully connected layers. The training dataset is kept up to date including 17,302 images of melanoma and nevus; this is the largest dataset by far. The model performance is compared to that of 157 dermatologists from 12 university hospitals in Germany based on MClass-D dataset. The model outperformed all 157 dermatologists and achieved state-of-the-art performance with AUC at 94.4% with sensitivity of 85.0% and specificity of 95.0% using a prediction threshold of 0.5 on the MClass-D dataset of 100 dermoscopic images. Moreover, a threshold of 0.40858 showed the most balanced measure compared to other researches, and is promisingly application to medical diagnosis, with sensitivity of 90.0% and specificity of 93.8%.