Research On Detection Of Micropapillary Lung Adenocarcinoma Based On Pathology Images

Micropapillary Lung Adenocarcinoma(MLA)is a histological subtype of lung adenocarcinoma.The micropapillary component is small and detached from the alveolar wall.Compared with the primary tumor,this irregular structure is more prone to metastasis and spread,showing high-intensity aggressive behavior,and is a high-grade malignant subtype with poor prognosis.Micropapillary pattern generally indicates the relatively advanced stage of development of lung adenocarcinoma,which is an extremely severe challenge for both patient and their family.If this histological type can be detected in time at the early clinical diagnosis,it has great clinical value for the early treatment of the disease.Histopathological image analysis is the gold standard for the diagnosis of micropapillary lung adenocarcinoma.At present,with the rapid development of artificial intelligence technology,artificial intelligence assisted pathological diagnosis has received extensive attention.However,due to the limited capabilities of deep learning models and the complexity of histopathological images,the existing computer-aided diagnosis still has the following two problems,which need to be further solved:(1)Insufficient label in pathological images,and the available training data is less.The training of neural network requires a large number of labeled data,but the size of the whole slide image(WSI)is huge,and only specialist pathologists can perform region-level annotations.Manual annotation is time-consuming and labor-intensive,and the cost of labeling is high,which leads to few histopathology image dataset with annotations.(2)Pathological image object multiscale problem.Digital pathology images have complex texture and high resolution,both the background information at the bottom and the detailed information at the top of the image are very important.However,the traditional single-scale models usually have problems such as insufficient information acquisition of small window and the increase of model parameters caused by large window.Therefore,it is necessary to integrate the information from multiple scales and design the detection model according to the analyzed object.Based on the above problems,the thesis aims to optimize the detection scheme of micropapillary to assist early clinical diagnosis and focus on the two problems of insufficient labeled data and multi-scale detection in pathological images.The main contributions and innovations of the thesis include the following two aspects:(1)In this thesis,a semi supervised learning framework is proposed to assist in the detection of micropapillary lung adenocarcinoma pathological images with limited annotations,to solve the problem of insufficient label in pathological images.The learning framework consists of two steps:fully supervised training and teacher-student semi-supervised training.The fully supervised training is trained with manually annotated data,and a one stage detection model composed of three parts,Backbone,Neck,and Head is designed.Complete-Io U(CIo U)loss is introduced for prediction of bounding box regression,combined with the efficient Non-Maximum Suppression(NMS)method Cluster-NMS to select the optimal target box,and suppress redundant regression boxes to complete the post-processing.In the teacher-student semi-supervised training stage,first,the model obtained in the fully supervised training is used as the teacher model to generate predictions for the unlabeled data.Secondly,confidence filtering is carried out on the prediction data,and the data whose predicted score is high than the threshold is converted into pseudo labels,while introducing various transformations of unlabeled data,jointly train to generate student models.Finally,iterate the teacher-student to train new student model with new pseudo labels.Extensive experiments were tested on the whole slide micropapillary lung adenocarcinoma histopathology image dataset,the detection precision and recall reached 0.775 and 0.896 respectively.The experimental results verify that the model can effectively utilize the unlabeled data to achieve rapid and accurate detection.(2)In this thesis,a multiscale model with circle representation is designed to optimize the detection of micropapillary in pathological object,to solve the problem of multiscale object in pathological images.It takes the model proposed by the Visual Geometry Group(VGG)of Oxford University as the feature extraction framework,and integrates the feature pyramid network(FPN)into the model to obtain the information such as texture and color in the low-level feature map,the cell shape and other information in the high-level feature map.When calculating the intersection over Union(Io U)between the generated candidate box and the ground truth,according to the characteristic contour of the micropapillary structure of the analyzed object,the circle Io U is designed as an evaluation metric,and the circle representation is used to replace the rectangular detection box.The redundant box is suppressed by Non-Maximum Suppression to output the detection results finally.After experiments on the micropapillary lung adenocarcinoma histopathology image dataset,the model achieves precision of 0.741 and recall of 0.852.The experimental results prove that the model realizes the optimization for the detection of micropapillary in pathological tissue.Based on the research work of this thesis,the computer-aided detection method designed in this thesis will contribute to the early recognition and diagnosis of micropapillary in clinic and lay a foundation for the research of micropapillary detection algorithm in pathological images.