Research On Image Enhancement And Segmentation Network Based On Mammography Lesion Distribution Information

Breast cancer is one of the most common cancers affecting women,with an estimated 30% of all new female cancer diagnoses attributed to this condition.Fortunately,the introduction of mammography for screening and diagnosis has been associated with a reduced risk of death from breast cancer,ranging from 28-36%.Computer-aided diagnosis has become one of the essential and important tools in the medical field due to the booming technology.Medical image enhancement and deep learning segmentation networks as two of the important auxiliary tools can greatly reduce the time and labor cost of diagnosis.Clinical studies and statistics have also found that breast tumors typically occur in the upper outer quadrant of the breast,indicating that these lesions have a distributional nature.Despite the progress made in computer-aided diagnosis,further work is needed to fully consider lesion characteristics and improve the accuracy of auxiliary diagnosis.This thesis will focus on image enhancement and semantic segmentation based on the distribution characteristics of breast lesion data,with the aim of achieving more accurate results.(1)A region-based mammography data enhancement algorithm is proposed in this thesis.We propose a region-adaptive image enhancement algorithm based on the statistics of lesions from public datasets.Based on the statistical results of pathologies in the open dataset,we propose an area-adaptive image enhancement algorithm to supplement the deficiencies of existing histogram equalization algorithms in local enhancement.This algorithm improves the clipping threshold in CLAHE by establishing the mapping relationship between lesion accumulation distribution and clipping threshold,reasonably distributing the pixels outside the threshold,and thus improving the contrast of mammography image regionally.Finally,this thesis evaluated multiple object detection,semantic segmentation and classification task models on two datasets,and the experimental results proved the effectiveness of this augmentation algorithm on object detection and classification tasks.It also affirmed the enhancement ability of the algorithm at the level of image vision and histogram.Although this enhancement algorithm has limitations in segmentation tasks,the lesion distribution characteristics it uses also provide a basis for the subsequent model establishment.(2)This thesis proposes a segmentation model based on lesion region and scale information perception module.In order to fully exploit the regional and multi-scale characteristics of lesion information,we propose an end-to-end U-shaped information perception segmentation neural network IPNN,extending the exploration of mammography data characteristics in segmentation networks.IPNN aims to perceive the distribution and scale information of lesions and promote the network to learn the fused features.First,the graph convolutional module based on regional information perception is responsible for dynamically learning the features between regions of the breast,and improving the feature expression of regional information.Furthermore,we add the Atrous Spatial Pyramid Pooling module in multiple skip connections to enhance the perception of scale information.Finally,the comparison experiments and ablation experiments on the public datasets prove that the two information perception modules proposed in this thesis can effectively improve the performance of the segmentation network,and the visualization results also demonstrate the superiority of the model.In summary,based on the lesion distribution characteristics of mammography data,this thesis proposes a novel regional image enhancement algorithm and a semantic segmentation network with an information perception module to address the existing problems of enhancement algorithms and segmentation networks.The related results demonstrate the efficacy of the proposed algorithm and components.