Development Of Intelligent Identification Platform For Circulating Tumor Cells Based On Convolutional Neural Network

The identification of Circulating Tumor Cells(CTC)smears is mainly achieved by the manual operation by pathologists.This method is limited by the subjective judgment of pathologists,so the diversity of the pathologists’ technical level and experience will directly affect the objectivity and accuracy of the results,and manual testing is time consuming,laborious,and inefficient.However,as the demand for CTC identification continues to increase.Therefore,the highly-automatic,accurate CTC intelligent identification platform is of high significance.Based on the functional requirements of the platform,this paper integrates the multidisciplinary knowledge of mechanical,biological,computer and control science,and conducts overall design and experimentation to achieve three major functions: Fully autonomous scanning path planning,automatic,and secondary inspection assistance.This platform contains a software system and automatic microscope hardware equipment.The hardware is composed of a microscope,micro-scanning system,CCD camera,etc.The software system is mainly composed of human-computer interaction interface and corresponding algorithms.The micro-scanning system adopts high-precision screw drive form to ensure the accuracy of the moving position.On this basis,a full autonomous scanning path planning method for CTC smear is proposed.We focused on designing a multi-feature fusion convolutional neural network architecture to identify three types of images inside,outside,and outside the smear.In order to overcome various difficulties caused by human or environmental factors in the segmentation process of CTC,such as uneven illumination and staining,cell adhesion,environmental impurities,etc.This paper proposes a deep cavity residual convolutional neural network integrated with multi-scale context information for segmentation detection of CTC images,and use the dense conditional random field method post-process the preliminary results,achieving 84.1% JA semantic segmentation accuracy and calculate relevant cell parameters to provide more accurate index data for the pathologist’s reexamination.Besides,this paper designs the rechecking auxiliary module,for example,using the location-based method to stitch CTC smear image quickly,a cluster ant colony algorithm is proposed to efficiently search the shortest traversal path,and using visual control to place the detected CTC in the central view,etc.Thus,a complete CTC smear intelligent identification process can be realized.We design human-computer interaction interface and integrate these proposed algorithms into it,besides,multi-threading technology is used to realize the coordination work between several modules such as micro-scanning platform control,image acquisition,and CTC segmentation.Finally,the actual test was carried out on the identification platform.The results show that the average detection time of a circular filter CTC smear with a diameter of 17 mm is 1.8 hours,the average accuracy is 86.69% and the average recall rate is 95.17%,thus the developed platform meets the design requirements.In summary,the CTC intelligent identification platform developed in this paper achieves efficient,objective and accurate detection of CTC,which can better meet the requirements of pathologists for CTC detection.