Study Of Method Based On Modality-Correlation Learning For Breast Tumor Diagnosis

Breast cancer is a disease with a relatively high mortality rate in women.Early diagnosis and treatment are the key to curing breast cancer.Medical imaging is the main auxiliary tool for the early diagnosis of breast cancer,where molybdenum target and ultrasound are the most commonly used imaging in breast cancer clinical examination.Multimodal image fusion techniques can overcome the shortcomings of insufficient single modal information and have become a research focus in recent years.However,most of the traditional methods are based on single mode classification first and then fuse the classification results.Such methods ignore the use of multimodal correlation information and limiting the performance improvement.In order to improve the performance of the breast tumor classification system,this paper conducts research on the key technologies of multimodal breast images fusion classificationIn this paper,a breast tumor classification method based on multimodal breast image correlation embedding(Modality-Correlation Embedding Model,MCM)is studied.First,the FCN network was used to segment tumor regions in multimodal breast images and the VGG network was used for tumor features extraction.Then,a new classification loss function of multimodal breast image fusion based on modality-correlation embedding was constructed.This loss function consists of a new modal correlation term and two single modal data fitting term.Finally,using gradient descent method and alternating variables iterative optimization idea to optimize the proposed loss function,two optimal mapping matrices are learned.Therefore,the depth features based on the molybdenum target image and the ultrasound image are more accurately mapped to the common label space,and the final diagnosis result is obtained.In the method of this paper,the proposed the modal correlation term can effectively make use of the correlation information between the multimodal images,keeping the features of different modalities in the label space tight,thereby ensuring that the different modalities of homologous patients remain consistent in the label space.In order to verify the effectiveness of the method,a relatively systematic experiment was carried out on a homologous multimodal data set containing 73 patients(42 cases were benign tumors and 31 cases were malignant tumors,and each patient respectively provided a molybdenum target image and an ultrasound image).In this paper,the method proposed reached 95.83%,95%,91.67%,95.83%,95.83%and 88.89%on AUC,accuracy,sensitivity,specificity,PPV and NPV,respectively.Its performance is generally better than traditional multimodal breast image fusion classification methods.The work in this paper further improves the accuracy of breast cancer diagnosis,and also enriches the multi-modal medical image fusion diagnosis technology to a certain extent,which has certain clinical value and application prospects.