Neural Network Based CAD System for the Classification of Textures in Liver Ultrasound Images

Radiologists interpret and classify the texture on the liver ultrasound images as normal or abnormal based on the discrepancies in the appearance of the lesion and the echogenicity. In this research, many textural metrics are evaluated from the region inside the lesion as well as from the normal region to automate the manual interpretation carried out by the radiologist. Following that, ten distinct neural networks are trained using these parameters. The initial weights for the neurons in each of these ten neural networks are distinct, and the training samples within the training dataset are distributed randomly. The results from each of these 10 neural networks are integrated in order to raise the classification precision and the confidence level of the system to classify the lesions. The experimental simulation results on liver ultrasound images with focal liver lesions (cyst, hemangioma, hepatocellular carcinoma, metastasis) and on normal liver images shows an improvement in the classification accuracy (93.89%) in comparison to 10 neural networks (2.17–8.33% improvement). Additionally, experimental findings demonstrate that, in comparison to individual classifiers, the proposed system classifies several classes with a high degree of confidence and the confidence level has increased by 12–44%, according to the results.