Automated Diagnosis of Melanoma Histopathological Images Based on Deep Learning Using Trust Counting Method

Improving the speed and accuracy of pathological diagnosis is a key challenge for pathological diagnosis and treatment instructions. And the automated diagnosis of melanoma histopathology slides is one of the potential methods in this task. We proposed a deep learning-based automated diagnosis system focusing on automatic diagnosis, which can diagnose melanoma in histopathological images end-to-end. In the system, an image processing algorithm was used for image slicing and image preprocessing, a ResNet50 model was used for patch-level prediction, and a novel statistics method called the trust counting method was proposed for slide-level diagnosis. The system's derivation and validation were implemented in a multi-centre database, including 701 WSIs from multiple body parts. In the melanoma pathological slices task, the system achieved an area under the receiver operating characteristic curve of 0.986, a sensitivity of 0.938, and a specificity of 0.947. And compared with the classical statistical methods, the system has achieved higher performance. The results of our study suggest that the proposed deep learning diagnosis system has feasibility for melanoma diagnosis.