Research On Breast Mass Detection And Classification With MRI Using Region-based CNN

Breast cancer is the second-highest incidence of cancer in the world,and it is the cancer of highest incidence in women.Imaging examination plays an important role in the early detection and diagnosis of breast cancer.With its precise and non-invasive way,it can detect early breast disease and improving the cure rate.Magnetic Resonance Imaging(MRI),as an important non-invasive way for cancer detection,can detect early breast lesions,accurately reflect the morphological characteristics and hemodynamic information of the lesions,and is an important tool for early detection of breast cancer.At the same time,with the continuous expansion of deep learning in the image field,how to combine deep learning and medical imaging technology to build computer-aided diagnosis algorithms,relieve doctors’ pressure on image reading,and improve the accuracy and stability of diagnosis,has become a problems in the field that need to be solved urgently.In order to solve the above problems,the main research work of this paper is to build a computer-aided detection and diagnosis algorithm for early detection and diagnosis of breast cancer masses on breast MRI images based on the region-based convolutional neural network.Due to the lack of public breast MRI datasets,this paper first processed and annotated the breast MRI data of the partner hospital to produced a standard dataset for breast MRI mass detection.Then improved the classic region-based convolutional neural network algorithm Faster RCNN to achieve highly sensitive detection of breast masses on the standard dataset.Finally,to solve the problems of difficult detection of small-size masses and low detection accuracy of benign masses,this paper integrate the feature pyramid network and cascade strategy in the Faster RCNN algorithm,it’s called Cascaded Feature Pyramid Network,which further improves the detection and classification performance of breast masses.Specifically,the main research work of this paper is as follows:(1)Make a standard dataset for breast MRI mass detection.For the original data of 191 patients included in the dataset,first using wavelet denoising to denoise these breast MRI images,and then labeling all masses under the guidance of doctors based on the diagnosis records,and this paper proposed a three-channel RGB image fusion method in data augmentation step,to obtain a breast RGB fusion image with stronger characterization ability and increase the sample space of the dataset.Finally,five consecutive slices were sampled from the front and back of the patient’s slice with largest tumor for experiments.(2)Modified the classic algorithm Faster RCNN to achieve high-sensitivity detection of breast masses.In this paper,Res Net101 is used as the backbone network of Faster RCNN,and a suitable anchor size is designed according to the size of masses,which improves the feature extraction ability and detection performance of the algorithm.In terms of loss function,Focal Loss is used to replace the original cross-entropy loss,so that the algorithm improves the attention to difficult samples and the balance between easy and hard samples.In the model training stage,a two-step transfer learning method is proposed,which makes the network learn priori knowledge from the natural image datasets Image Net and PASCAL VOC respectively.Finally,our algorithm reaches 0.837 m AP on the breast mass dataset,and 0.979 sensitivity(0.401 FPs/Volume).(3)The feature pyramid network and cascaded strategy are integrated into the Faster RCNN algorithm to construct a Cascaded Feature Pyramid Network,which further improves the algorithm’s performance in detecting and classifying breast tumors.Due to the small masses are easily missed and the detection rate of benign masses is much lower than that of malignant masses during the experiment,this paper proposed a multi-scale feature pyramid structure to improve the backbone network of Res Net101,which improving the network’s ability to recognize small masses;A cascade strategy is adopted to cascade the single-threshold network into a multi-threshold network to improve the network’s learning ability for benign masses.Finally,the m AP for detecting breast masses was increased to 0.860,and the detection performance for small masses was greatly improved.In summary,all the work in this paper is focused on the task of detecting and diagnosing masses in breast MRI images,which provides a method reference for the subsequent application.