Axel Wismueller MD, Presenter: Nothing to Disclose

Gerda Leinsinger MD, Abstract Co-Author: Nothing to Disclose

Anke Meyer-Baese PhD, Abstract Co-Author: Nothing to Disclose

Raffaela Wiegard, Abstract Co-Author: Nothing to Disclose

Oliver Lange PhD, Abstract Co-Author: Nothing to Disclose

Maximilian Reiser MD, Abstract Co-Author: Nothing to Disclose

To develop, test, and evaluate a neural network machine learning approach for characterization of diagnostically challenging breast lesions in MRI.

88 women with 92 indeterminate mammographic lesions (BIRADS III-IV, 41 benign and 51 malignant lesions confirmed by histopathology, median lesion diameter 12 mm) were examined by standardized dynamic contrast-enhanced breast MRI on a 1.5 T system. The machine learning approach is based on a 3-layer feed-forward radial basis functions (RBF) neural network for automatic classification of the 6-dimensional lesion-specific signal intensity (SI) time course vectors. As a reference, interactive visual classification of SI time courses was performed by human experts (initial SI increase and post-initial SI time course) according to a standardized semi-quantiative evaluation score (Kuhl et al., Radiology 211:101-110). In addition, morphological criteria were included based on a clinically approved scoring system. Quantitative evaluation of classification performance was obtained by a meticulous 25-run multiple leave-out cross-validation setup, where 50 data sets were randomly selected for neural network training, and the remaining 38 data sets for test in each run. For quantitative assessment of diagnostic accuracy, areas under ROC curves (AUC) were computed for both machine learning and human expert classification.

The neural network machine learning approach increased both sensitivity and specificity for classification between benign and malignant breast lesions, as confirmed by quantitative analysis of diagnostic accuracy: Neural network results (AUC=80±4%) clearly outperformed human expert evaluation of SI time-series with (AUC=64±5%) and without (AUC=59±6%) consideration of lesion morphology. The increase in diagnostic accuracy for neural network classification proved to be statistically significant (Wilcoxon matched-pairs, two-sided, p<0.05).

Automatic lesion classification in breast MRI by RBF neural networks is a powerful and cost-effective method for computer-aided diagnosis in suspicious lesions leading to a substantial improvement of diagnostic accuracy beyond the visual interactive classification by human experts.

Wismueller, A, Leinsinger, G, Meyer-Baese, A, Wiegard, R, Lange, O, Reiser, M, Computer-aided Diagnosis in Dynamic Breast MRI: Can Machine Learning by Neural Networks Improve the Diagnostic Accuracy in Suspicious Lesions?.  Radiological Society of North America 2004 Scientific Assembly and Annual Meeting, November 28 - December 3, 2004 ,Chicago IL. http://archive.rsna.org/2004/4414608.html