Digital Mammography-based Microcalcification Radiomics Signature For Breast Cancer Classification And Prediciton

Microcalcifications(MCs),a frequent mammographic sign of precancerous cells or early breast cancer,pose serious challenges to clinicians for accurate diagnosis due to their inappreciable variation in distribution and morphology between malignancy and benignity.Highly suspicious non-palpable MCs in radiology inevitably requires invasive core needle biopsy examinations in pathology as the gold standard.Therefore,in order to reduce unnecessary biopsy pains and to provide clinical guidance via exploiting the massively underlying information within non-invasive medical images,this paper established a relatively complete paradigm of MCs-focused and task-specific radiomics study for breast cancer in the following three aspects.Firstly,shape properties and texture patterns coupled with multiscale image transformation techniques were explored as distributional descriptors for MCs.A series of local statistical features of pixel intensity,i.e.,energy,entropy and gray-level co-occurrence matrix(GLCM),from derivative transformed patches such as wavelet decompositions were extracted.Subsequently,in terms of individual MCs specks,this paper constructed their topological structures using vertices and edges in graph theory as representation across a spectrum of scales.Interpretable topological metrics have been generated to explore different pathological status of MCs by their connectivity,density,linearity and compactness.In the end,artificial intelligence(AI)-aided radiomics approach was performed that included data standardization,feature dimensionality reduction through one-factor statistic tests and classifier-embedded algorithms successively and classifications based on machine learning methods.Simultaneously,task-specific performance assessments of the proposed discriminative model were also taken into account.Positive Predictive Value(PPV)and Sensitivity(SE)served as malignancy-related metrics,alongside Negative Predictive Value(NPV)and Specificity(SP)as benignity-related.The overall performance was measured by the Area under the Curve(AUC)with its Receiver Operating Characteristic curve(ROC)analysis and Accuracy(ACC).