Construction And Application Of Contrast Enhanced Mammography Based Artificial Intelligence Diagnosis Model Of Breast Cancer

Breast cancer is the most common malignant tumor in women and its incidence is increasing year by year.It is also a leading cause of cancer-related deaths.Early detection,diagnosis,and treatment are crucial in reducing breast cancer mortality rates.The progressive use of contrast-enhanced mammography(CEM)and artificial intelligence(AI)in medical imaging plays a significant role in accurately diagnosing breast cancer.CEM involves injecting an iodine contrast agent to enhance the visibility of breast lesions,allowing for clearer visualization.It can provide information on the morphological characteristics and blood supply of the lesions.AI technology can assist doctors in early diagnosis and screening of breast cancer by analyzing and learning from a large number of images.This study aims to build different breast cancer diagnostic models based on CEM imaging features,radiomics(including intratumoral and peritumoral)characteristics and clinical information.Machine learning methods will be employed to build and develop these models.Finally,the application value of different models in clinical practice was evaluated by a multi-reader reading study.This research is mainly divided into the following three parts.1.Machine learning breast cancer diagnosis model based on CEM BI-RADS classification image featuresPurpose:To develop a machine learning model that can accurately distinguish between benign and malignant breast masses using CEM BI-RADS classification image features,including low-energy image features and recombined image features.Additionally,an interpretability module will be utilized to investigate the impact of each image feature on the model’s classification.The study will also explore how common artifacts present in CEM images affect the accuracy of BI-RADS classification by diagnosticians.Materials and Methods:A retrospective collection was conducted,including a.total of 244 patients with suspicious breast masses who underwent CEM examination at Nanfeng Hospital between March 2018 and December 2021.These patients were subsequently confirmed through pathological biopsy.Experienced radiologists extracted the imaging features of low-energy and recombined image based on the BI-RADS classification system.Using histopathology as the gold standard,an XGBoost machine learning model was constructed,and its diagnostic performance was evaluated using the area under curve of receiver operating characteristic curve(AUC of ROC)as the evaluation metric.To analyze the influence of image features on model performance,SHapley Additive explanations(SHAP)values were utilized.Additionally,the study recorded the BI-RADS classification and common artifacts observed in CEM images,and further analyzed the impact of these artifacts on BI-RADS classification accuracy.Results:The machine learning breast cancer prediction model based on CEM BI-RADS image features,achieved an AUC of 0.961.The sensitivity,specificity,and accuracy of the model were 0.909,0.867,and 0.892,respectively.Analysis of SHAP values revealed that specific image features,such as " circumscribed or obscured mass margin","trabecular thickening",and "calcifications absent" in low-energy images,as well as " lesion conspicuity high" features in recombined images,along with the demographic information of "age",significantly influenced the predictive performance of the model.Most artifacts observed in CEM recombined images did not affect the BI-RADS classification diagnosis.However,for benign breast lesions,the presence of vascular enhancement that impacted the lesion led to a higher likelihood of diagnostic physicians classifying the lesion as a BI-RADS category of 4B or higher.Conclusion:The machine learning model based on CEM BI-RADS classification image features is helpful for the diagnosis of breast cancer.The morphological and enhancement characteristics of the mass play a crucial role in distinguishing between benign and malignant lesions.In the case of benign breast lesions,the presence of a vascular enhancement artifact causes diagnostic physicians to classify the lesion into a higher BI-RADS category.2.Construction of the intratumoral and peritumoral radiomics model based on CEM for prediction breast cancerPurpose:Development of breast cancer prediction radiomic models using intratumoral and peritumoral features extracted from CEM images.Explore which area around the lesion can provide the most valuable diagnostic information.Materials and Methods:A retrospective analysis was performed on 228 patients who received CEM in Nanfeng Hospital from March 2018 to December 2021.All lesions were histopathologically confirmed.The region of interest(ROI)was manually delineated,while the algorithm automatically obtained the annular region of interest around the lesion.Four ROIs were obtained for each lesion,including the lesion ROI,annular lesion margin ROI(-1mm to 1mm),and annular lesion peripherality ROIs(3mm,5mm).Radiomics features within each ROI were extracted after image preprocessing.The dataset was randomly divided into a training set and a test set in a 3:1 ratio.Different classification models were constructed using XGBoost with the extracted radiomics features.The performance of each model was evaluated using AUC(95%CI),sensitivity,and specificity.Result:Among 228 breast lesions,93 were benign and 135 were malignant.In the test set,the lesion ROI based model achieved the highest AUC of 0.918(95%CI:0.844,0.991),with a sensitivity of 85.9%,specificity of 90.5%,and accuracy of 87.9%.Regarding the ROI models based on the surrounding area of the lesion,the model utilizing the lesion edge-1mm to 1mm region had the highest AUC of 0.791(95%CI:0.665,0.917).In the models combining the features of the lesion ROI and the ROI of the surrounding area,the fusion model of lesion+lesion edge-1mm to 1mm area demonstrated the highest AUC of 0.833(95%CI:0.721,0.944).Conclusion:The lesion ROI-based radiomics model exhibits good discriminatory ability between benign and malignant breast lesions.However,further investigation is needed to determine the value of radiomics features from the surrounding areas in breast lesion diagnosis.3.Clinical application of artificial intelligence classification model of breast lesions based on CEM images--image reading testPurpose:To evaluate the diagnostic value of two artificial intelligence breast cancer prediction models for radiologists.Materials and Methods:Five radiologists with different experience in breast diagnosis were selected to read CEM images of 119 patients independently,and BI-RDS classification was performed for each lesion with or without radiomics model assistance respectively.Additionally,the radiologists extracted image features and inputted them into the first part of a machine learning model to assess its prediction efficiency.The performance of each physician was evaluated using AUC,sensitivity,specificity,and accuracy.Results:The average AUC of the radiologists using the conventional method was 0.863(95%CI:0.736-0.939),while the average AUC of the machine learning prediction model was 0.848(95%CI:0.770-0.891).The image-based machine learning models improved the diagnostic specificity for radiologists with 3-5 years of experience(Reader 2:40.05%vs,61.9%;Reader 3:52.4%vs.69.1%;Reader 4:45.2%vs.61.9%).The average AUC of the radiologists in the diagnosis with radiomics model assistant was 0.950(0.915-0.989),which was improved by different experienced radiologists.Notably,Reader 1(1 year experience)demonstrated the most significant improvement(0.736 vs.0.937,P<0.05).With the model assistance,BI-RADS 4A and BI-RADS 4B were significantly reduced.Conclusion:The machine learning model based on image features can improve the specificity and accuracy of diagnosis for radiologists with intermediate experience,while it did not show significant improvement for those with low or high experience levels.On the other hand,the model based on radiomics can improve the diagnostic performance of radiologists with different experience,particularly in terms of specificity,thus demonstrating wider applicability.to previous studies.