Radiomics Study Of Nonpalpable Breast Lesions With Microcalcification-only On Mammography

Part I The study of texture features of DWI differentiation of malignant from benign nonpalpable breast lesion on patients with microcalcifications-only in mammographyObjectiveUsing texture feature of DWI differentiation of malignant from benign nonpalpable breast lesion on patients with Breast Imaging-Reporting and Data System(BI-RADS)3-5 microcalcifications-only in mammography.MethodsThis is a retrospective study approved by the Internal Research Review and Ethical Committee of the Zhejiang Cancer Hospital.We enrolled 61 patients from October 2012 to December 2015 of Zhejiang Cancer hospital in patients with BI-RADS 3?5 microcalcifications-only on mammography(postoperative of the three-dimensional X-ray wire localization of breast,the pathology was confirmed,including 38 patients with malignant lesions and 23 patients with benign lesions),whose DWI images were analyzed.All images segmentations were performed by two expert board-certified breast radiologists who had more than 20 years of experience,and they were blinded to the pathological results.According to the location of the lesions at the maximum diameter of the MRI axial enhanced image,the lesions on DWI were aligned and identified,and the areas of interest were delineated using the ITK-SNAP software(http://www.itksnap.org/pmwiki/pmwiki.php).Within each ROI,we extracted 6 histogram features and 16 grayscale symbiosis matrix(GLCM)texture features.The classification of benign and malignant lesions of postoperative pathology was used as the standard to train the classifier.The random forest algorithm was hired to select the features and build the classification model,leave one out cross validation(LOOCV)was used to validate the classifier,the performance of the classifier was evaluated by AUC.ResultsThe texture feature combination with the best classification performance on DWI contains 6 features,including the histogram features of mean,variance,skewness,entropy and GLCM features of 0°direction of contrast,45°direction of correlation.The AUC of the model reached 0.76,and the diagnostic accuracy,sensitivity and specificity were 77.05%,84.21%and 65.21%respectively.The histogram features of mean,variance,skewness and entropy were significantly different between the benign and malignant groups.ConclusionThe texture feature analysis of DWI can improve the diagnostic accuracy of differentiating benign and malignant breast nonpalpable lesions with microcalcifications-only on mammography to some extent.Variance,skewness and entropy are new imaging marker with great potential.Part ? The study of multimodality radiomics differentiation of malignant from benign nonpalpable breast lesion on patients with microcalcifications-only in mammographyObjectiveTo investigate a radiomic scheme that combines image features from digital mammography and dynamic contrast-enhanced magnetic resonance imaging to improve classification performance on nonpalpable breast lesion(NBL)with Breast Imaging-Reporting and Data System(BI-RADS)3-5 microcalcifications-only in mammography.MethodsThis is a retrospective study approved by the Internal Research Review and Ethical Committee of the Zhejiang Cancer Hospital.The study consisted of 81 patients who had BI-RADS 3-5 microcalcifications-only in mammography between October 2012 to November 2016 consecutively(postoperative of the three-dimensional X-ray wire localization of breast,the pathology was confirmed,including 40 patients with malignant lesions and 41 patients with benign lesions),who had pretreatment breast MRI and mammography both.All images segmentations were performed by two expert board-certified breast radiologists who had more than 20 years of experience separately,and they were blinded to the pathological results.They delineated the suspicious malignant calcifications on mammography and enhanced lesions on DCE-MRI images as the areas of interest by using of rTK-SNAP software(http:/www.itksnap.org/pmwiki/pmwiki.php)Within each ROI,we extracted morphological features,histogram features and texture features from the two modalities.In addition,we extracted features based on the specific diagnostic value of each modality.According to the postoperative pathological report,40 cases of malignant tumors were included in the malignant group and 41 cases of benign lesions were included in the benign group.The classification of benign and malignant lesions of postoperative pathology was used as the standard to train tthe classifier.We used the random forest recursive feature elimination method to reduce the features on each single modality to obtain the combination of features with the best performance and then combined them.We used the random forest algorithm to train the multimodal classifier to differentiate the malignant NBL from the benign ones.And we validated using the Leave-One-Out-Cross-Validation method and assessed using an area under the receiver operating characteristic curve to evaluate its discriminating performance.ResultsFinally,106 candidate radiomics features were reduced to 14 potential features from two modalities.The multimodal classifier achieved AUC of 0.903,with a sensitivity of 82.5%,a specificity of 80.48%,was better than any single modality.ConclusionMultimodal radiomics classification shows promising power in discriminating malignant lesions from benign lesions on NBL patients with BI-RADS 3-5 microcalcifications-only in mammography.This research may provide a more accurate and objective noninvasive way before making an invasive clinical strategy.