Analysis of liver scintigrams by neural network

Colloid liver scintigraphy has been performed widely for diagnosis of diffuse parenchymal liver disease. Pattern of these disease such as chronic hepatitis or liver cirrhosis is evaluated by size and distortion of the liver, distribution of tracer in the liver, size and activities of tracer in the spleen, visualization of the bone marrow and so on. It is not difficult to read a scintigram which shows typical pattern of normal, chronic hepatitis and liver cirrhosis, but in some cases it is difficult to read normal or chronic hepatitis and chronic hepatitis or liver cirrhosis in visual diagnosis. So we tried to use a neural network to make differential diagnosis accurately. First, five features in colloid liver scintigrams were evaluated visually. These features were left lobe/right lobe length splenomegaly, visualization of the bone marrow, liver deformity and distribution of tracer in the liver. Analysis of liver scintigrams was carried out by the neural network using these features. True positive ratio in the neural network was as follows. Normal : 92% (23/25); Chronic hepatitis : 71% (27/38) ; Liver cirrhosis : 93% (39/42). The relationship of diffuse parenchymal liver disease and these features also was found by analysis of the inner structure in the neural network. Thus one concludes that the neural network is useful on colloid liver scintigraphy because differential diagnosis can be done with 85% in overall true positive ratio, and the method is simple and useful clinically. (author)