基于部分监督学习的通用病变检测方法OA

This paper proposes a new loss function calculation strategy to reduce the possibility of misclassifying unlabeled le-sion areas into normal tissues during training.Before calculating the loss function,the model adopts a new automatic selection mech-anism of unlabeled samples to realize the automatic selection of negative samples.At the same time,in the process of loss calcula-tion,the negative sample adaptive attenuation coefficient is set to reduce the proportion of unlabeled samples,which could further reduce the adverse impact of misclassification on the model.The method proposed in this paper is trained on the current optimal model.The data set uses the open partial annotation dataset DeepLesion,which contains the partial annotation data of CT images from eight different organs of the human body.A large number of experiments have been carried out to select the most appropriate learning rate parameters to make the model effect reach the optimal state.The results show that when the FPPI values are 0.5 and 1,the sensitivity is improved by 1.1％and 0.92％,respectively,compared with the state-of-the-art general lesion detection model.Applying the loss function calculation strategy proposed in this paper to the model can effectively improve the performance of the general lesion detector.