抄録
The purpose of this study is to compare four CAD-algorithms under the same condition, and to verify the most effective algorithm for the classification of microcalcifications on mammogram. The four CAD-algorithms are BP-ANN, GA-ANN, FL and GA-FL. BP-ANN is a conventional artificial neural network with back-propagation learning. GA-ANN is an improved ANN based on genetic algorithm to determine the weighting coefficients at ANN. FL is a conventional fuzzy logic algorithm using gaussian-distributed membership functions. GA-FL is an improved FL based on genetic algorithm to optimize the membership functions. Comparison results are indicated by ROC curves. The Az values of the ROC curves for BP-ANN, GA-ANN, FL and GA-FL were 0.86, 0.80, 0.89 and 0.95, respectively. When sensitivity of each algorithm was 100%, specificities of BP-ANN, GA-ANN, FL and GA-FL were 69%, 54%, 31% and 77%, respectively. Therefore, these results show the effectiveness of GA-FL algorithm. GA-FL has a potential to become useful CAD algorithm to classify microcalcifications on mammogram.
