Artificial Neural Networks for differential diagnosis of breast lesions in MR-Mammography: A systematic approach addressing the influence of network architecture on diagnostic performance using a large clinical database

作者：Dietzel Matthias^*; Baltzer Pascal A T; Dietzel Andreas; Zoubi Ramy; Groeschel Tobias; Burmeister Hartmut P; Bogdan Martin; Kaiser Werner A

来源：European Journal of Radiology, 2012, 81(7): 1508-1513.

DOI：10.1016/j.ejrad.2011.03.024

• Neoplasms-Primary
• Artificial Neural Networks
• Breast
• Magnetic resonance imaging
• Pathology
• Diagnostic imaging
• Computer aided diagnosis
• Lesion characterization

Rationale and objectives: Differential diagnosis of lesions in MR-Mammography (MRM) remains a complex task. The aim of this MRM study was to design and to test robustness of Artificial Neural Network architectures to predict malignancy using a large clinical database. %26lt;br%26gt;Materials and methods: For this IRB-approved investigation standardized protocols and study design were applied (T1w-FLASH; 0.1 mmol/kgBW Gd-DTPA; T2w-TSE; histological verification after MRM). All lesions were evaluated by two experienced (%26gt;500 MRM) radiologists in consensus. In every lesion, 18 previously published descriptors were assessed and documented in the database. %26lt;br%26gt;An Artificial Neural Network (ANN) was developed to process this database (The-MathWorks/Inc., feed-forward-architecture/resilient back-propagation-algorithm). All 18 descriptors were set as input variables, whereas histological results (malignant vs. benign) was defined as classification variable. Initially, the ANN was optimized in terms of %26quot;Training Epochs%26quot; (TE), %26quot;Hidden Layers%26quot; (HL), %26quot;Learning Rate%26quot; (LR) and %26quot;Neurons%26quot; (N). Robustness of the ANN was addressed by repeated evaluation cycles (n: 9) with receiver operating characteristics (ROC) analysis of the results applying 4-fold Cross Validation. The best network architecture was identified comparing the corresponding Area under the ROC curve (AUC). %26lt;br%26gt;Results: Histopathology revealed 436 benign and 648 malignant lesions. Enhancing the level of complexity could not increase diagnostic accuracy of the network (P: n.s.). The optimized ANN architecture (TE: 20, HL: 1, N: 5, LR: 1.2) was accurate (mean-AUC 0.888; P: %26lt;0.001) and robust (CI: 0.885-0.892; range: 0.880-0.898). %26lt;br%26gt;Conclusion: The optimized neural network showed robust performance and high diagnostic accuracy for prediction of malignancy on unknown data.

• 出版日期2012-7