Using neural networks to aid the diagnosis of breast implant rupture

From a database consisting of 78 implants that were surgically removed, ultrasound findings and surgical results were used to train and test backpropagation and radial basis function (RBF) neural networks using the round-robin or leave-one-out method. Receiver-operating-characteristic curve analysis was applied to compare the performance of the different neural networks with that of the radiologists involved in the ultrasound evaluations. The neural networks outperformed the radiologists involved. RBF networks performed better in this classification problem than did backpropagation networks. The best performing network utilized, in addition to the findings, the (unaided) diagnosis of the radiologist. Thus, the ‘team’ approach appears to provide the best results. Also, the network performed particularly well in those cases in which the radiologist classified the implant as indeterminate. The results suggest that a neural network using findings extracted from sonograms by experienced sonographers can be of great assistance to physicians with the diagnosis of implant rupture.