Research On Detection Algorithm Of Breast Mass In Mammogram

Mammography screening is an important technique for detecting breast cancer in early stage,of which the key step is to detect breast masses in mammography images accurately.However,due to the heterogeneity of the breast mass and the complexity of its surrounding environment,it has become a challenging problem to develop a robust detection method for the breast mass.After analyzing the existing breast mass detection technology,a breast mass detection scheme based on deep learning algorithm was designed.In the image preprocessing module,the adaptive effective grayscale normalization method is applied,combining with the contrast limited adaptive histogram equalization algorithm to improve the contrast between the breast mass and the surrounding environment,clearer the mass boundary.Moreover,the natural deformation-based data augmentation solution is used to solve the problem of overfitting caused by the limitation of data set size.By simulating the growth pattern of breast mass,the complexity and diversity of data set are effectively enhanced.In the breast mass detection module,the main architecture is implemented based on a one-stage target detection model Retina Net with FSAF module and feature pyramid network,and the corresponding optimization is performed according to the information distribution characteristics of mammography image.Due to the existence of feature pyramid network,this architecture can well mitigate the impact occasioned by the different breast mass sizes in the raw data.At the same time,in order to make full use of limited data,the detecting difficult level of samples in the verification set is used as a benchmark,during the training process,samples with high difficult level in the verification set are exchanged dynamically with the random samples of the training set,so as to effectively mine the information amount of the difficult samples in the verification set.Finally,in order to further improve robustness of the detecting model,we referring to the idea of transfer learning to assist model training.Ablation studies and comparison experiments with existing related methods are carried out based on the public datasets INbreast and DDSM,which proved the effectiveness of strategies such as image preprocessing,data augmentation,network structure improvement,and dynamic data exchange in this paper.On INbreast dataset,when each image has an average of 0.495 false positives,the recall rate is 0.930;Whilst on DDSM dataset,when each image has an average of 0.599 false positives,the recall rate is 0.943,which can accurately detect breast masses.