We report on the derivation and implementation of the first three-dimensional optical tomographic image reconstruction scheme that is based on the time-independent equation of radiative transfer (ERT) and allows for arbitrarily shaped medium boundaries and arbitrary spatial material distributions. The scheme builds on the concept of model-based iterative image reconstruction, in which a forward model provides prediction of detector readings, and a gradient-based updating scheme minimizes an appropriately defined objective function. The forward model is solved by using an even-parity formulation of the ERT, which lends itself to a finite-element discretization method. The finite-element technique provides the suitable framework for predicting light propagation in arbitrarily shaped three-dimensional media. For an efficient way of calculating the gradient of the objective function we have implemented an adjoint differentiation scheme. Initial reconstruction results using synthetic data from simple media and a three-dimensional mesh of the human forehead illustrate the performance of the code. © 2003 SPIE and IS&T.