Computational assessment of breast tumour differentiation using multimodal data

Early detection of breast cancer requires accurate prediction and reliable diagnostic modalities. This allows physicians to distinguish malignant tumours before proceeding with a painful surgical biopsy. The attributes of three non-invasive primary diagnosing modalities, namely symptomatic examination, ultrasound imaging, and mammographic results, were used for the study. A dataset was created using ten selected features from each modality, after iterations during the training phase. The number of satisfying features was used for the creation of a model, which was further categorised as benign or malignant class. The model was evaluated in the testing phase by comparing biopsy results for benign or malignant classification. The statistical analysis proved it as an efficient approach for non-invasive decision-making. The developed model was tested using supervised learning algorithms with three classifiers for 210 cases by comparing the results with the gold standard biopsy results. The sensitivities for the three classifiers were 80%, 73% and 76.5%, while specificities were 96%, 94.4% and 95%, respectively. This method of breast tumour differentiation using the features of the non-invasive modalities can be widely used in telemedicine applications, helping to reduce confirmatory biopsies.