Classification of breast disease with MRI using neural networks

Introduction: Due to overlap in the appearance of benign and malignant lesions on post-contrast scanning, dynamic contrast enhanced MR imaging (DCE-MRI) has been used to aid differentiation. This study investigates one Artificial Intelligence technique. neural networks, to analyse DCE-MRI, compared to two radiologists, one experienced in MR Breast imaging, the other to a lesser extent. Methods: DCE-MRI was performed on 103 patients (69 malignant, 34 benign). Twenty-five images were acquired with a temporal resolution of 11.6 sec. Regions of interest were drawn around the entire abnormality and time-course data for each pixel presented to a probabilistic neural network (PNN). The radiologists then examined both the time course data and the post-contrast images separately and then in combination. Results: Analysing the time course graphs alone the PNN achieved the best classification with an accuracy of 90%, compared with 76% and 56% respectively for the radiologists. Utilising the postcontrast images alone the accuracy of the radiologists increased to 89% and 85% respectively. With the post-contrast images and time course data combined the radiologists accuracy increased to 97% and 93% respectively. Conclusions: The PNN provides an accuracy similar to that of an experienced radiologist, who utilised post-contrast images alone, and could prove beneficial to radiologists with limited MR experience of breast imaging.