Indecisive Trees for Classification and Prediction of Knee Osteoarthritis

Random forests are widely used for classification and regression tasks in medical image analysis. Each tree in the forest contains binary decision nodes that choose whether a sample should be passed to one of two child nodes. We demonstrate that replacing this with something less decisive, where some samples may go to both child nodes, can improve performance for both individual trees and whole forests. Introducing a soft decision at each node means that a sample may propagate to multiple leaves. The tree output should thus be a weighted sum of the individual leaf values. We show how the leaves can be optimised to improve performance and how backpropagation can be used to optimise the parameters of the decision functions at each node. Finally, we show that the new method outperforms an equivalent random forest on a disease classification and prediction task.