Neural network analysis of breast cancer from mammographic evaluation

Mammographic differentiation of benign lesions from malignancies is a difficult task. We developed an artificial neural network [ANN] as a diagnostic aid in mammography using radiographic features as input. A three-layered ANN was used to differentiate malignant from benign findings in a group of patients with proven breast lesions on the basis of morphological data extracted from conventional mammograms. Our database included 122 patient records on 14 qualitative variables. The database was randomly divided into training and validation samples including 82 and 40 patient records, respectively, to construct the ANN and validate its performance. Sensitivity, specificity, accuracy and receiver operating characteristic curve [ROC] analysis for this method and the radiologist were compared. Our results showed that the neural network model was able to correctly classify 30 out of 40 cases presented in the validation sample. Comparing the output with that of the radiologist, showed a reasonable diagnostic accuracy [75%], a moderate specificity [64%] and a relatively high sensitivity [89%]. A diagnostic aid was developed that accurately differentiates malignant from benign pattern using radiological features extracted from mammograms

Improving the performance of neural network in differentiation of breast tumors using wavelet transformation on dynamic MRI

A computer aided diagnosis system was established using the wavelet transform and neural network to differentiate malignant from benign in a group of patients with histo-pathologically proved breast lesions based on the data derived independently from time-intensity profile. The performance of the artificial neural network [ANN] was evaluated using a database with 105 patients' records each of which consisted of 8 quantitative parameters mostly derived from time-intensity profile using wavelet transform. These findings were encoded as features for a three-layered neural network to predict the outcome of biopsy. The network was trained and tested using the jackknife method and its performance was then compared to that of the radiologists in terms of sensitivity, specificity and accuracy using receiver operating characteristic curve [ROC] analysis. The network was able to classify correctly the 84 original cases and yielded a comparable diagnostic accuracy [80%], compared to that of the radiologist [85%] by performing a constructive association between extracted quantitative data and corresponding pathological results [r=0.63, p<0.001]. An ANN supported by wavelet transform can be trained to differentiate malignant from benign breast tumors with a reasonable degree of accuracy