A deep learning fusion model with evidence-based confidence level analysis for differentiation of malignant and benign breast tumors using dynamic contrast enhanced MRI

•Fused model of deep learning and manually extracted image features of breast tumor.•Temporospatial discriminative image features provided evidence of classification.•100% certainty can be achieved considering the evidence-based confidence criteria. Accurate interpretation of breast magnetic resonance imaging (MRI) remains challenging due to the lack of highly-specialized experiences of radiologists. The purpose of this study is to develop a Convolutional Neural Network (CNN) model based on dynamic contrast enhanced (DCE) MRI images as a reliable evidence-based computer aided diagnostic tool for breast cancer diagnosis to augment diagnostic performance. This retrospective study included a total of 130 patients (71 malignant and 59 benign tumors). Surgical removal or biopsy procedures were performed for pathological analysis as the ground truth of tumor malignancy. The CNN model fused analytical features of tumor geometric information and pharmacokinetic properties of tumor tissues for differentiation of malignant and benign breast tumors. The confidence level of the CNN model classification outcome was interpreted based on three types of evidence criteria: prediction probability, feature hotmap visualization, and contributive dynamic scan time points. The CNN model achieved an overall diagnostic accuracy of 87.7%, precision of 91.2%, sensitivity of 86.1%, and AUC of 91.2% ± 4.0% across five-fold testing process. The evidence-based model confidence analysis precluded misclassification given that all three types of evidence were met. This CNN-empowered computer aided diagnostic tool with evidence-based confidence level analysis achieved high diagnostic performance in breast tumor classification based on a single DCE MRI sequence. This work indicated that it is feasible to shorten breast cancer MRI screening protocol by including one DCE sequence, as well as to augment radiologist’s efficiency by focusing on cases with lower model confidence.