A neural network made of a Kohonen's SOM coupled to a MLP trained via backpropagation for the diagnosis of malignant breast cancer from digital mammograms

Abstract

A system was built entirely based on artificial neural networks to be used as an aiding tool in the analysis of mammograms for the diagnosis of breast cancer. The system receives a mammogram as input and gives as output one of three possible answers: suspicious of malignant breast cancer, suspicious of benign breast cancer, and without suspicion of breast cancer. The system uses two different kinds of neural networks: Kohonen's self-organizing map and multilayer perceptron trained with the backpropagation algorithm. The system's performance was evaluated by its ability to generalize after training. The true-positive fraction or sensitivity (fraction of actually malignant cases that was correctly classified) was 0.50, and the false-positive fraction (fraction of actually benign or normal cases that was incorrectly classified) was 0.12. On the other hand, the fraction of actually malignant or benign cases that was correctly classified was 0.75 and the fraction of normal cases that was incorrectly classified either as benign or malignant was 0.38.