Research On Cells Detection Algorithm Of ThinPrep Cervical Pathological Images

Thinprep cytologic test(TCT)is an important method for cervical cancer screening.Object detection algorithms can assist doctors in locating and classify diseased cells in TCT slides.However,the existing algorithms’ performance is limited due to missing labels in the training data,the class imbalance of the training data,and a large number of cells in the slides with different characteristics.In order to solve these problems,this paper puts forward the corresponding solutions,which can be summarized as follows:(1)To solve the problem of missing labels in network training,we proposed a TCT object detection network training method for missing labels.This method improves the object detection network’s existing training process and integrates the module named label correction network into the object detection network.According to the set rules,the label correction network corrects the training samples’ labels during the training process.we conducted experiments on TCT datasets,and natural image datasets proved that our approach could effectively improve the performance of the object detection model on the training data containing missing labels.(2)To solve the problem that the performance of diseased cells’ detection is limited due to the class imbalance of training data,we propose a TCT object detection network training method against class imbalance.On the one hand,the data augmentation proposed in this thesis,which can be used to improve the attention of the network to the features of pathological cells in the training process.On the other hand,the model’s prediction robustness on minority classes of diseased cells is enhanced by optimizing the loss function.The experiment proved that our method could effectively improve the object detection network’s performance on TCT images.(3)To solve the incorrect classification of slides caused by false-positive detection results,we proposed a cascade network-based TCT negative slide screening algorithm.The cascade network is composed of two parts: detection network and fine-grained feature mining network.The fine-grained feature mining network learns the false-positive results output features by the detection network and filters out the false-positive results output by the detection network when analyzing the slices.Finally,according to the number of diseased cells detected in the slides,finished the positive and negative classification of the slices.The experimental results show that the proposed algorithm’s specificity can reach 27.3% when the algorithm’s sensitivity is 100%..To sum up,this paper analyzes the factors that restrict object detection algorithms’ performance in practical application and puts forward the corresponding solutions.It is committed to helping to improve the diagnostic accuracy of pathologists while reducing the workload of doctors,to promote the smoothly for our country cervical cancer screening work.