Micro-Calcification Clusters Detection Method Based On Deep Learning

Deep learning is a neural network model with extensive multi-level structure proposed in recent years.Its earlier applied research is mainly aimed at the field of image recognition.With the rapid development of current artificial intelligence technology,big data and AI applications are driving the transformation of digital life into intelligent life,intelligent computing systems will help people handle everyday tasks and get useful information more quickly and easily.The existing deep learning algorithm model has surpassed humans in some areas of lesion detection for medical images,computer-aided detection is effective in reducing misdiagnosis or missed diagnosis during examination.Microcalcification clusters are early indicators of the onset of breast cancer,and their induced lesions are the primary factors that endanger women’s health.Because the detection rate of microcalcification clusters is very low,accurate identification and localization are important for the early diagnosis of breast cancer.(1)In order to obtain the effective feature expression of microcalcification cluster lesions in X-ray images,this paper proposes a FEC-Net(Fine-grained cascade enhanced)convolutional neural network model based on WRI(Wide-Residual-Inception)convolution framework.The extraction of lesion features by FCE-Net has been improved by the multi-level structure of the neural network and the various types of structure of the nucleus neurons,the residual connection structure introduces element-level weight accumulation between levels to enhance the sensitivity of the change of the feature image output,so as to solve the problem of gradient disappearancecaused by network depth.For the detailed features of microcalcification clusters,the multi-branch feature variation structure enhances the width of the neural network and solves the over-fitting problem caused by excessive network parameters.(2)Microcalcification clusters are one of the lesion categories in X-ray images.Effective classification and regionalization are the core of target detection.Due to the small granularity of small target detection,a candidate detection network model based on Multi-scale Fusion of Features(MFF)is proposed.Each pixel on the depth profile of the microcalcification cluster based on FCE-Net corresponds to the local receptive field on the original image,including microcalcification clusters and other lesions.By constructing the MFF candidate detection network,the multi-scale features are multi-sampled and multi-scale,and the parallel calcification and feature classification are obtained,and the micro-calcification cluster confidence and regional coordinates are obtained.Finally,the target region is subjected to label classification and boundary regression.(3)The FCE-Net+MFF deep neural network model is verified on the MIAS dataset.By comparing the efficiency difference caused by the minimization loss function in the model training,the influence of the training error and the generalization error is concentrated,and the test result is fed back and the parameter setting of the model is adjusted.The experimental results show that the effective algorithm optimization makes the detection efficiency of the neural network model effectively improved.The microcalcification cluster detection method proposed by the research institute has certain practicability,and can play a computer-aided detection role in medical image detection and cancer prevention,and further explores the target detection algorithm combined with deep learning in future research work.