Prediction Of Breast Cancer Metastasis Based On Multi-Objective Optimization Under Multi-Modal Data Fusion

Breast cancer(BC)is the most common malignant tumor in women and the second leading cause of cancer death in women.The distant metastasis of BC,causing failure of other organs is a common cause of death caused by BC.Therefore,the use of artificial intelligence technology to predict distant metastasis of BC and early detection of the risk of distant metastasis of BC can help patients to take countermeasures in advance,thereby effectively reducing the risk of death of BC patients and improving the prognosis.This paper first improved the multi-objective genetic algorithm NSGA2 and proposed the MINSGA2 algorithm for the feature selection and classification model optimization.Based on the idea of mutual information,the algorithm improves three aspects of the multi-objective genetic algorithm: population initialization strategy,crossover mutation operator and selection operator.That is,the population initialization strategy based on label correlation,the local search operator based on relevant and redundant values,called supplement and delete operator,and the information score of feature subsets.The ablation experiment was carried out on the three improvement points,which proved that our model got the excellent improvement effect.After that,this paper used multiple public data sets for feature selection and classification prediction,which proved that the algorithm can achieve good classification results on different data sets.A single-modal experiment for breast cancer distant metastasis prediction was conducted on structured data modalities using clinical datasets,and an excellent single-modal classifier was obtained.The AUC of the model reached 0.931,the sensitivity was 0.945,and the specificity was 0.917.In addition,we conducted a null hypothesis experiment on the body composition data,which proved that body composition data played a positive role in correctly predicting distant metastasis in BC patients.Afterwards,this paper studied the image modality of breast cancer.According to previous studies,there is a great correlation between the body composition of BC patients and their prognosis.Therefore,this paper used CT images of the fourth thoracic vertebra(T4)and eleventh thoracic vertebra(T11)that can prompt the body composition information of BC patients,used deep learning methods to predict distant metastasis of BC.Besides,we used the idea of deep radiomics,to extract image features and prepare for subsequent multi-modal fusion research.Based on the idea of channel attention,this paper combined the channel attention in SENet and ECANet in a parallel manner to propose a mixed channel attention module called MCI module,and integrated it into multiple alternative networks.Finally,a well performed network Dense Net121-MCI was obtained,with an AUC of 0.825,a sensitivity of 0.796,and a specificity of 0.854.Based on this,this paper obtained a single-modal classifier for distant metastasis of breast cancer under the image data modality,and obtains image data extraction features.Finally,this paper conducted a variety of multi-modal fusion experiments based on the ideas of early fusion,late fusion,and hybrid fusion.In the part of early fusion,the features of the two modalities were fused,the principal component analysis method was used to reduce feature redundancy,and the MINSGA2 method was used for BC metastasis prediction,which achieved the excellent results.In the part of later fusion,the single-modal classifiers obtained were used for fusion experiments.Finally the multi-objective weight optimization method achieved the excellent performance.Therefore,in the hybrid fusion experiment,we used the NSGA2 for weight optimization hybrid fusion,and finally obtained the BC distant metastasis multi-modal fusion classifier.The AUC reached 0.980,the sensitivity was 0.987,and the specificity was 0.975.In addition,this paper used the idea of Grad-CAM to draw hot-spot images for the features extracted by deep learning,and finds that the attention area of network features is concentrated on the patient’s erector spinae muscle area,which coincides with previous experiments and proves again that body composition positive effect of data on prediction of distant metastasis in BC patients.