Neural network-based approach for the non-invasive diagnosis and classification of hepatotropic viral disease

A neural network-based method is presented for the diagnosis and classification of patients infected with hepatotropic viral disease. The non-invasive approach makes use of real-time pathological data in form of liver function tests and specific virological markers during training. The unknown patients, not included in the training set, are then processed through the trained neural network model. Experimental results demonstrate that the proposed method is able to effectively diagnose the disease and classifies its stage to be acute, chronic or cirrhosis. In addition, performance analysis is carried out for various supervised and unsupervised neural network models with optimal topologies. The topological architectures are chosen after experimenting with various training algorithms using diverse parameters such as different activation functions between layers, number of hidden layers, and number of neurons. It is concluded that the performance of the multilayered supervised feedforward neural network is the most accurate even with small data set. Whereas, the unsupervised Kohonen´s self-organising maps do not perform well for the subject task under similar conditions.