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.

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. 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 (>500 MRM) radiologists in consensus. In every lesion, 18 previously published descriptors were assessed and documented in the database. 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 "Training Epochs" (TE), "Hidden Layers" (HL), "Learning Rate" (LR) and "Neurons" (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). 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: <0.001) and robust (CI: 0.885-0.892; range: 0.880-0.898). CONCLUSION: The optimized neural network showed robust performance and high diagnostic accuracy for prediction of malignancy on unknown data.

Application of breast MRI for prediction of lymph node metastases - systematic approach using 17 individual descriptors and a dedicated decision tree.

BACKGROUND: The presence of lymph node metastases (LNMs) is one of the most important prognostic factors in breast cancer. PURPOSE: To correlate a detailed catalog of 17 descriptors in breast MRI (bMRI) with the presence of LNMs and to identify useful combinations of such descriptors for the prediction of LNMs using a dedicated decision tree. MATERIAL AND METHODS: A standardized protocol and study design was applied in this IRB-approved study (T1-weighted FLASH; 0.1 mmol/kg body weight Gd-DTPA; T2-weighted TSE; histological verification after bMRI). Two experienced radiologists performed prospective evaluation of the previously acquired examination in consensus. In every lesion 17 previously published descriptors were assessed. Subgroups of primary breast cancers with (N+: 97) and without LNM were created (N-: 253). The prevalence and diagnostic accuracy of each descriptor were correlated with the presence of LNM (chi-square test; diagnostic odds ratio/DOR). To identify useful combinations of descriptors for the prediction of LNM a chi-squared automatic interaction detection (CHAID) decision tree was applied. RESULTS: Seven of 17 descriptors were significantly associated with LNMs. The most accurate were "Skin thickening" (P < 0.001; DOR 5.9) and "Internal enhancement" (P < 0.001; DOR

Application of artificial neural networks for the prediction of lymph node metastases to the ipsilateral axilla - initial experience in 194 patients using magnetic resonance mammography.

BACKGROUND: In breast MRI (bMRI), prediction of lymph node metastases (N+) on the basis of dynamic and morphologic descriptors of breast cancers remains a complex task. PURPOSE: To predict N+ using an artificial neural network (ANN) on the basis of 17 previously published descriptors of breast lesions in bMRI. MATERIAL AND METHODS: Standardized protocol and study design were applied in this study (T1w-FLASH; 0.1 mmol/kg body weight Gd-DTPA; T2w-TSE; histological verification after bMRI). All lesions were evaluated by two experienced radiologists in consensus. In every lesion 17 previously published descriptors were assessed. Matched subgroups with (N+; n=97) and without N+ were created (N-; n=97), forming the dataset of this study (n=194). An ANN was constructed ("Multilayer Perceptron"; training: "Batch"; activation function of hidden/output layer: "Hyperbolic Tangent"/"Softmax") to predict N+ using all descriptors in combination on a randomly chosen training sample (n=123/194) and optimized on the corresponding test sample (n=71/194) using dedicated software. The discrimination power of this ANN was quantified by area under the curve (AUC) comparison (vs AUC=0.5). Training and testing cycles were repeated 20 times to quantify the robustness of the ANN (median-AUC; confidence intervals, CIs). RESULTS: The ANN demonstrated highly significant discrimination power to classify N+ vs N- (P<0.001). Diagnostic accuracy reached "moderate" AUC (median-AUC=0.74; CI 0.70-0.76). CONCLUSION: Application of ANNs for the prediction of lymph node metastases in breast MRI is feasible. Future studies should evaluate the clinical impact of the presented model.

Magnetic resonance mammography of invasive lobular versus ductal carcinoma: systematic comparison of 811 patients reveals high diagnostic accuracy irrespective of typing.

OBJECTIVE: Invasive lobular (ILC) and ductal carcinomas (IDC) are the most frequent subtypes of breast cancer. Diagnosis of ILC is often challenging. This study was conducted to (1) evaluate dynamic and morphologic profiles and to (2) compare the diagnostic accuracy of IDC and ILC in magnetic resonance mammography (MRM). METHODS: Our database consisted of all consecutive MRMs over a 12-year period (standardized protocol: T1-weighted fast low-angle shot; 0.1-mmol gadolinium-diethylenetriaminepentaacetate per kilogram of body weight; T2-weighted turbo spin-echo, 1.5 T; histological verification after MRM), which were evaluated by experienced (>500 MRMs) radiologists in consensus, applying 17 predefined descriptors. All the patients gave written consent; this study was approved by the local institutional review board. Extracting all the ILCs (n = 108), IDCs (n = 347), and benign lesions (n = 436) from the database, the data set of the study was created.In ILC and IDC diagnostic accuracy of single descriptors was calculated and compared separately (chi test). Using all the descriptors, a combined binary logistic regression analysis was applied to calculate the overall diagnostic accuracy for ILC and IDC. The corresponding areas under the curve were compared. RESULTS: ILC and IDC, showed wash-in and an irregular shape without difference (P = 1.0 and P = 0.4). Wash-out was more typical of IDC (72.6%; ILC, 57.4%; P = 0.007). Perifocal edema was diagnosed more frequently in IDC (45.5%; P = 0.05). For overall accuracy, the areas under the curve were 0.929 for ILC and 0.939 for IDC (P = 0.5). CONCLUSIONS: The dynamic and morphologic profiles of ILC and IDC were overlapping, and minor differences between both subgroups could be identified. Accordingly, the overall diagnostic accuracy of MRM was high and without difference between both subtypes of breast cancer.

Differential diagnosis of breast lesions 5 mm or less: is there a role for magnetic resonance imaging?

OBJECTIVES: According to magnetic resonance (MR) imaging Breast Imaging Reporting and Data System, foci are small enhanced lesions 5 mm or less in diameter. This study was conducted to (a) assess morphological and dynamic profiles in malignant versus benign foci in breast MR imaging (MRM) and to (b) identify overall diagnostic accuracy of MRM for differential diagnosis of foci. METHODS: This study was approved by the local institutional review board; all patients gave written consent. All MRM (T1w-FLASH; 0.1 mmol/kg body weight gadolinium-diethylenetriamine penta-acetic acid; T2w-TSE; consecutive 12-year period; with histological verification after MRM were evaluated by 2 experienced (>500 MRM) radiologists in consensus using 16 predefined descriptors and were included into a database. A data set was created by extracting all lesions 5 mm or less (benign, 27; malignant, 61). Accuracy of individual descriptors was assessed (Crosstabs, chi2-test; positive/negative likelihood ratios (LR+/-); diagnostic odds ratio [DOR]). Binary logistic regression analysis was applied to identify overall diagnostic accuracy using all descriptors combined (area under the receiver operating characteristic curve,). RESULTS: Washout was typically associated with malignancy (P < 0.05; DOR, 3.5). Irregular shape was feasible for differential diagnosis of foci (DOR, 7.3), yet majority of malignancies demonstrated a round shape (55.6%). Additional descriptors such as blooming (DOR, 4.0, LR+, 2.8), adjacent vessel (DOR, 4.8; LR+, 4.5), and root sign (DOR, 5.6; LR+, 4.1) showed a high accuracy. Overall accuracy for differentiation of benign versus malignant foci showed an area under the curve of 0.887 (P = 0.0001). CONCLUSIONS: Assessment of dynamic and morphological profiles in foci 5 mm or less was feasible. Using all descriptors combined, a high potential for differential diagnosis of foci in magnetic resonance-mammography could be identified.