Algorithm Research And System Implementationof The Nucleus Detection Based On Fully Convolutional Network

Pathological diagnosis is regarded as the "gold standard" for cancer diagnosis.It can not only clarify the origin of cancer tissue,but also provide accurate classification and typing of cancer tissue,providing important reference and guidance for clinical treatment.As an important basis for pathological diagnosis,by observing its morphology,size,number and arrangement,pathologists are able to accurately determine the type and extent of disease,and thus formulate precise prognostic strategies.Therefore,the detection of nucleus is of irreplaceable significance in the field of pathology.With the continuous development of artificial intelligence technology and digital pathology,the use of automated technology to detect nucleus in digital pathology images has become a research hotspot.However,due to the heterogeneity of the nucleus itself,the imbalance of foreground and background samples,and the overlapping and adhesion of nucleus,the existing automatic nucleus detection algorithms have low precision and many false positive results.In response to these problems,this thesis proposes a neural network algorithm model named DBFCN for the first time.In order to further improve the detection performance of this model,this thesis also designs a dual-branch dynamic interaction mechanism based on prior position information.At present,the DB-FCN algorithm and model have been experimentally verified on multiple nucleus detection datasets,and the experimental results have fully confirmed its highprecision advantages.In addition,to meet the actual clinical needs,this thesis also builds an AI digital pathological image nucleus detection system based on the DB-FCN algorithm model we proposed.The system realizes automatic nucleus detection and visual display of digital pathological images,which reduces the work pressure of pathologists to a certain extent and improves the efficiency and accuracy of pathological diagnosis.The main work and innovation of this thesis are summarized as follows:1.In view of the situation that the one-stage detection algorithm is efficient but low in accuracy,and the two-stage detection algorithm is high in accuracy but low in efficiency,this thesis proposes a new end-to-end dual-branch fully convolutional network model(DB-FCN),which successfully combining the advantages of one-stage and two-stage detection algorithms.In this algorithm,the coarse detection branch will transfer the prior location information to the fine detection branch,so that the fine detection branch can learn under the given confidence of the nucleus of each position,so as to avoid interference from noise and background.Therefore,the fine detection branch can focus on suppressing the false positive results of the coarse detection output,so that the model can achieve better accuracy.2.Based on the DB-FCN model,this thesis designs a dual-branch dynamic interaction mechanism based on prior location information.Through the dynamic interaction of the two detection branches,the detection performance of the model can be further improved and false positive results can be reduced.At present,the algorithm has been tested on three public nucleus detection datasets of CRCHisto Phenotypes,Pannuke and cc RCC,and the experimental results have fully confirmed the effectiveness of the mechanism.3.To meet the actual clinical needs,this thesis developed an AI digital pathological image nucleus detection system.The system can detect nucleus in digital pathology images quickly,accurately and reliably.Compared with traditional nucleus detection systems,the system can not only perform nucleus detection on Patch images cropped from digital pathology images,but also realise complete nucleus detection and visualisation functions for digital pathology images.The system brings great convenience to clinical applications and provides better diagnosis and treatment decision support for doctors.