Abstract
The renography has been of an increasing interest in the diagnosis of the renal function. But it is difficult to deduce the renal function from the quantitative estimation of the renograms because of many factors involved. For the practical diagnosis it has been common to classify renograms by observation into several types. The pattern recognition by observation is not accurate ; for the same renogram, sometimes different identification is given by different physicians, even by the same doctor at different times.
This paper describes an algorithm for pattern recognition of renograms with a small digital computer by applying the so-called “Learning Machine” method. The renogram-patterns were classified into 7 types. Of the 7 types, 2 were separated from the others by simple logic. To the other 5 types, the “Learning Machine” method was applied. In our method, 10 linear discriminant functions must be computed to establish the decision surfaces. This enabled our “Learning Machine” to classify unknown renograms into one of 5 renogram types. Moreover, the machine is devised so as to be able to get more appropriate decision surfaces by learning new renograms one by one.
On the first learning, the machine learned 25 renograms, which consisted of 5 typical renograms per each 5 types. When the first learning was finished, the percentage of agreement between machine and specialist classifications was 58 %. When the number of renograms learnt by the machine was increased to 53, it became 90 %. This fact proves that our algorithm is clinically applicable enough.
