Research On Tumor Histopathological Image Analysis Method Based On Deep Semi-Supervised Learning

Histopathological images are crucial in diagnosing malignant tumors and determining treatment plans.The conventional diagnosis method is to diagnose through the visual microscope of the pathologist,this process is time-consuming,labor-intensive,and subjective leading to varying results among pathologists.With the rapid development of computational pathology,there is an opportunity to achieve objective diagnoses,predict treatment reactions,and identify new morphological characteristics.However,to train neural networks in computational pathology,a large number of labeled datasets are required,which usually be costly and time-consuming to obtain.Semisupervised learning aims to make use of unlabeled data to train models with limited labeled data,reducing the burden of data labeling.Considering the practical clinical needs,this thesis aims to fully utilize both labeled and unlabeled data in tissue pathological image analysis by employing semi-supervised deep learning techniques.To achieve this goal,several analysis algorithms based on semi-supervised deep learning are proposed in this study.The main research contents of this thesis are as follows:(1)This thesis proposes a semi-supervised tumor pathological image analysis model based on consistency regularization to address the challenge of analyzing pathological images of tumor tissues such as breast cancer.The model leverages consistency regularization and a pseudo-tagging strategy to generate artificial tags for untagged images,effectively utilizing untagged data.In addition,the consistent regularization loss function is modified to generate empty tags for some images during training to remove irrelevant regions.This is important since whole slide images may contain a large number of such regions.According to the image features,the image processing conversion database is defined for data augmentation.(2)This thesis introduces a novel semi-supervised segmentation algorithm based on category content activation.This model generates an initial segmentation response using image-level labels as the supervision for the semantic segmentation model.The classification and segmentation networks are integrated into an end-to-end model to improve the quality of the segmentation response.A joint loss function is added to optimize both branches simultaneously.The algorithm generates a specific category corresponding activation region by distinguishing the activation layer.In addition,an additional conditional random field operation is performed on the activation area to modify it into a more reliable pseudo-label.This improves the accuracy of the segmentation response.(3)This thesis proposes a semi-supervised and weakly supervised segmentation algorithm based on consistency learning,which addresses the challenge of using various types of label data and unlabeled data in the segmentation task.Unlike existing semi-supervised segmentation methods that focus on using classification tags to improve class activation graphs,this thesis studies the teacher-student structure network and designs a semi-supervised segmentation model based on this structure.The model obtains the pixel probability distribution map from image-level annotation data and predicts pixel probability for untagged data.The confidence of pseudo-tags is defined to extract confidence regions,incomplete confidence regions,and disbelief regions from imagelevel labeling and untagged data,each corresponding to different loss functions.To enable the uncertain region to participate in model training,the thesis proposes using a contrast loss function to treat the uncertain region as a negative sample,thereby improving the segmentation effect.This thesis proposes a range of semi-supervised deep learning algorithms to address two medical problems: pathological image classification of breast cancer and tissue region segmentation of lung adenocarcinoma.The algorithms are designed to account for the specific characteristics of different tissue pathological images and meet the clinical analysis requirements.The semi-supervised tumor pathological image analysis algorithm based on consistency regularization enables model training with only a small amount of labeled data.This approach leverages the consistency of predictions across different samples to enhance model performance.The semi-supervised segmentation algorithm based on category content activation uses image-level label data to segment the target region,improving segmentation accuracy by focusing on regions of interest in the image.The semi-supervised and weakly supervised segmentation algorithms based on consistent learning use a combination of labeled and unlabeled data to train the segmentation model.This approach make full use of various types of label data and unlabeled data that may be obtained in the segmentation task to improve segmentation accuracy.The proposed pathological image analysis algorithm improves upon the semi-supervised deep learning method to address the challenge of limited labeled data in pathological image analysis.The algorithm has significant implications for the field of artificial intelligence,as it helps to realize intelligent pathological reading and improve the accuracy of diagnoses.