Designing a Hybrid Model to Differentiate between Malignant and Benign Patterns in Breast Cancer from Mammographic Findings

A genetic algorithm (GA) in conjunction with neural network was proposed to be used to differentiate between breast lesions based on the mass and microcalcification findings. These findings were encoded as features for a genetic algorithm for feature selection and classified with a three-layered neural network to predict the outcome of biopsy. The system was established through the optimization of the classification performance of neural network which was used as evaluation function. A database containing records of 119 patients (49 malignant and 70 benign cases) each of which consisted of 12 parameters was used. The artificial neural network (ANN) was trained using 79 cases containing masses and microcalcifications previously diagnosed by surgical biopsy and tested with 40 cases. The performance of hybrid model was then compared to that of the experienced radiologist in terms of sensitivity, specificity, accuracy and ANN was used to determine receiver operating characteristic curve analysis. The proposed approach was able to find an appropriate feature subset (5-7 parameters out of 12 parameters) and can approximately show the relative efficacy of each feature in classification of breast lesions. The optimized subset of features helped neural classifier to achieve better results in the classification of breast lesions. The obtained results showed that the neural network with optimized features yielded a higher diagnostic accuracy (80%), sensitivity (75%) and specificity (83%) compared to that of the radiologist (68%), (%83) and (61%). Similarly the output of the neural network with optimized features (hybrid model) outperformed the complete model by improving the accuracy, sensitivity and specificity from 70, 56 and 79% to 80, 75 and 83%, respectively. Discussion and The benefits of applying the GA as a preprocessor include improvement in the generalization ability of ANN and reducing the size of calculations through simplifying the ANN structure. Using such hybrid method demonstrated that the performance of the ANN improved by definition the importance of each evaluated features. This can lead the radiologist to a more accurate diagnosis in the clinical applications using such hybrid model as a computerized second opinion.