Research On Breast Image Mass Detection Algorithm Based On Cascade Deep Neural Network

In recent years,because of the rapid development of deep learning technology,target detection technology based on deep learning has been widely used in military,medical,agriculture,aviation,industry,and other fields.Breast cancer is considered to be one of the most leading cancers affecting women around the world.It has been a consensus that early detection with successful treatment played a crucial role to improve the survival rate.In clinical practice,mammography is a widely used diagnostic tool for detecting early symptoms of breast cancer.Most studies have shown that mammography can reduce breast cancer mortality.To correctly recognize troublesome areas and apply a correct diagnosis,radiologists face great challenges because they have to read a large number of breast images daily,and the reading process is monotonous,tiring,lengthy,costly,and most importantly,prone to errors.Otherwise,the shape,size,boundary,and texture of the masses are variety and the signal to noise ratio of the masses is lower compared to the surrounding breast tissue.Because their judgment depends on training,experience,and subjective criteria.Several studies have shown that 20-30% of diagnosed cancers have been missed in a retrospective study of previous mammograms.In order to alleviate the heavy work of radiologists and improve the sensitivity of detection in clinical practice,Computer-aided detection and diagnosis(CAD)are essential to provide physicians with “second” opinions to aid and support their decisions.In recent years,deep learning technology has gradually become the mainstream technology for the automatic detection of mammograms.The medical image analysis requires higher standards in accuracy for data set labeling,and the number of positive and negative samples are imbalanced.Serious imbalance can cause overfitting problems in the training model.There is diversity in the size,shape,border,and texture of the tumor in mammograms,and the signal-to-noise ratio of the tumor is lower than that of the surrounding tissues,the location of the tumor changes greatly,the appearance of the tumor changes greatly in different environments.There is little difference between the appearance of the target and the non-target.The above characteristics of mammography images bring great challenges to the accuracy of deep learning-based mass detection framework.Another challenge for current target detection algorithms is that many small size masses are missed during the detection process in mammograms.High-level features can better map to semantic information,while low-level features can better target.Expected results may occur if the two features are fused.During the training phase,CNN only extracts features from labeled candidate regions.However,due to the limited experience and cognition of pathologists,labeled areas are often subjective and even inaccurate.The cancer area depends on the area of residence which is difficult to define.That is,the output may not only depend on the input,but also on the region of the topology domain in which it is located.In response to the above issues,we present a novel approach for detecting masses in mammograms using a multi-stage object detection architecture with spatial constraints and multi-feature fusion,which is named SC-FU-CS RCNN.First,it consists of a sequence of detectors trained with increasing IoU thresholds,to be sequentially more selective against close false positives.Second,we improve the accuracy of the detection by concatenating the shallow and deep layers of the convolutional neural network(CNN),the detector can detect blurrier or smaller masses in mammograms.Finally,we add a spatial constrained layer before the output layer.This article consists of the following sections:(1)A background review of medical image analysis using computer-aided diagnosis.The development of computer-aided diagnosis based on deep learning target detection algorithms in recent years is described.(2)The advantages and disadvantages of Computer-aided detection methods based on hand-designed features and deep learning features are introduced.The two-stage and one-stage classical target detection algorithms are introduced respectively,including improvement methods,and current problems.(3)We present a novel approach for detecting masses in mammograms using a multi-stage object detection architecture with spatial constraints and multi-feature fusion,which is named SC-FU-CS RCNN.First,to solve the overfitting problem of the training model caused by sample imbalance.Using cascaded R-CNN target detection architectures,in which the detector is trained through multiple incremental overlap(IoU)thresholds to achieve the purpose of resampling samples.Second,we improve the accuracy of the detection by concatenating the shallow and deep layers of the CNN,the detector can detect blurrier or smaller masses in mammograms.Finally,a spatial constraint layer is added to incorporate the topological region features around the target into the training model.In order to improve the detection accuracy,the data set was enhanced in the experiment.(4)To effectively implement the multi-level target detection architecture based on spatial constraint layer and multi-feature fusion proposed in this paper,the published digital database for screening mammography(DDSM)and INbreast dataset were used for training.Quantitative evaluation of the SC-FU-CS RCNN method proposed in this paper is performed using evaluation criteria such as sensitivity,specificity,overall accuracy,and Roc curve.The performance of the improved Cascade R-CNN framework was compared with the other state-of-the-art deep-learning CAD systems.Besides,the experiments compared the detection performance of Improved Cascade R-CNN against the original ones.The results show that SC-FU-CS RCNN achieves higher accuracy than other algorithms.Our results provide feasible and promising results in terms of detecting the location of benign and malignant masses and recognize their proper classes as well.