Estimating the distributed diffusion coefficient of breast tissue in diffusion-weighted imaging using multilayer perceptrons

The aim of this article is to facilitate the estimation of the distributed diffusion coefficient (DDC) of breast tissue in quantitative diffusion-weighted imaging using artificial neural networks. Diffusion signals of breast lesions and of healthy glandular tissue of contralateral breasts from 174 women measured on diffusion-weighted images captured with a 3T MR scanner using small region of interests were used. Traditional DDC estimates were obtained by a stretched exponential model and nonlinear least-squares fitting applied to the diffusion signals. Various multilayer perceptrons having one hidden layer but with different number of neurons were developed. Diffusion signals normalized and DDCs estimated traditionally were the input vectors and the target outputs of the neural networks, respectively, that were randomly divided into training, validation and test datasets. Supervised leanings were performed with the training and the validation datasets using a backpropagation algorithm followed by tests with the test dataset. DDC estimation by least-squares fitting takes 38 ms, on average. Strong positive correlations are observable between the DDC estimates by least-squares fitting and by the neural networks (overall r = 0.962–0.999). However, a network having seventeen neurons in its hidden layer provides the strongest correlation (r = 1.000). Once the learning of the network is accomplished, the network computes a DDC estimate only in 73 μs without requiring any initial value or any boundary constraint. Multilayer perceptrons facilitate the estimation of the distributed diffusion coefficient of breast tissue in diffusion-weighted imaging by offering less computational complexity and reduced computation time compared to nonlinear least-squares fitting.