Research On Computer-Aided Diagnosis Key Algorithms Of Digital Breast Tomosynthesis Based On Deep Learning

Breast cancer has become the most common malignant tumor in women,posing a serious threat to their health.In China,there is still a large gap in the early diagnosis and survival rates of breast cancer patients compared to those in western countries.Promoting breast imaging examinations and achieving accurate early diagnosis of breast cancer are keys to improving patient survival rates.Digital Breast Tomosynthesis(DBT)is a three-dimensional breast X-ray imaging technology that can effectively reduce missed and misdiagnosed cases caused by overlapping glandular tissue and has an increasingly important role in breast cancer screening.However,the lesions in DBT are distributed in multiple layers,which significantly increases the workload of radiologists.In addition,the current early lesion diagnosis method based on DBT highly depends on clinical experience and lacks objective quantitative diagnostic criteria.The DBT breast cancer computer-aided diagnosis system based on deep learning algorithms can quickly and accurately provide objective quantitative diagnostic references for doctors,which has significant clinical application value.Therefore,this study focuses on the key issues in accurate diagnosis of early suspected lesions of breast cancer in DBT and conducts research on lesion detection and classification algorithms for microcalcification clusters and masses.The specific research contents and innovation points are as follows:(1)Research on microcalcification cluster false positive reduction algorithm based on anisotropic three-dimensional context-aware model.The false positive reduction of microcalcification clusters aims to distinguish true microcalcification clusters from false positive detection samples output by lesion initial detection algorithms to reduce the false positive rate of detection algorithms.To solve the problems of large intra-and inter-layer resolution disparity in DBT images,distribution of microcalcification clusters in multiple layers,and differences in sharpness,this study proposes an anisotropic 3D context-aware false-positive reduction algorithm.The method proposes to use anisotropic convolution-based intra-layer feature extraction and inter-layer contextual feature fusion operations,which ensures that cross-layer feature fusion occurs only at the feature map level,avoiding the effect of the anisotropic resolution of the original input DBT image.Due to the independence of intra-layer feature extraction,the method can also achieve feature reuse of the clearest slice(central slice)of microcalcification clusters without additional computing resource consumption.Experiments were conducted on the collected Chinese female breast dataset,and the algorithm achieved 92.68%classification accuracy and 0.9765 area under the curve(AUC)of the receiver operating characteristic(ROC)curve.The experiments showed that the algorithm can effectively use anisotropic 3D contextual information of DBT to improve classification performance.(2)Research on the benign and malignant classification algorithm of breast microcalcification clusters based on structure-aware hierarchical feature extractionTo address the problem of difficulty in effectively utilizing prior features of benign and malignant microcalcification clusters in DBT,this study proposes a structure-aware hierarchical feature extraction network for the classification of benign and malignant microcalcification clusters.This study first argues the principle defect of the 3D anisotropic convolution used in the previous study and proposes to use 2D grouped convolution for efficient hierarchical feature extraction,and demonstrates that grouped convolution can be considered a generalized form of anisotropic convolution.In addition,this study also proposes a microcalcification cluster structural feature extraction module based on the partial deformable Transformer,which effectively models the long-range dependency between microcalcification points,introducing the structural prior information of microcalcification clusters into the benign and malignant classification task.The experiment results on a private dataset showed that the algorithm achieved an average AUC of 0.8687,which is better than the algorithm used in the previous study and other state-of-the-art algorithms in the field.(3)Research on breast mass detection algorithm based on adaptive feature pyramid and uncertain boundary modelingTo address the problem of the difficulty in accurate detection of breast masses that have fuzzy boundaries and can be easily obscured by dense glandular tissue in DBT,this study proposes a detection algorithm based on an adaptive feature pyramid and uncertain boundary modeling.The adaptive feature pyramid module uses a crossattention structure to achieve direct global correlation modeling between multi-scale feature maps,avoiding the feature attenuation of traditional feature pyramid modules.The study also uses an uncertain boundary modeling-based bounding box representation method,which predicts a set of position vectors to discretely model the probability distribution of bounding box coordinates.This approach improves the accuracy of bounding box prediction while obtaining the uncertainty of bounding box coordinates and solves the model optimization problem caused by the uncertainty of the mass boundary location.Two independent experiments were conducted on both public and private datasets,and the algorithm obtained an average detection sensitivity of 90.76%and 92.86%respectively,demonstrating the value of the algorithm for DBT mass detection.In summary,this paper uses deep learning image analysis algorithms as the technical basis to carry out research on key issues such as accurate detection and classification of microcalcification clusters and masses in DBT,aiming to reduce false positives of microcalcification clusters,classify benign and malignant microcalcification clusters,and accurately detect masses.It provides a theoretical basis and technical support for strengthening the widespread application of deep learning in DBT image breast cancer computer-aided diagnosis technology.