In this paper, we address the stability issue of 2D ARMA models for ultrasound breast images, and use the estimated 2D ARMA coefficients as feature vectors for a multi-layer perceptron neural network, which classifies the ultrasound image into: (i) normal, (ii) benign tumor, or (iii) cancerous tumor. Furthermore, we test our CAD system using genetic neural learning algorithm. Specifically, we use the estimated 2D ARMA coefficients as inputs to a genetic neural network to classify the ultrasound breast image into three regions: healthy tissue, benign tumor, and cancerous tumor. Our simulation results on various cancerous and benign ultrasound breast images illustrate the power of the proposed algorithm.