Maximum likelihood (ML) estimation is used during tomosynthesis mammography reconstruction. A single reconstruction involves the processing of high-resolution projection images, which is both compute-intensive and time-consuming. This workload is presently a bottleneck in the accurate diagnosis of breast cancer during screening. This paper presents our parallelization work on an ML algorithm using three different partitioning models: no inter-communication, overlap with inter-communication and non-overlap model. These models are evaluated to obtain the best reconstruction performance given a range of computing environments with different computational power and network speed. Our test results show that the non-overlap method outperforms the other two methods on all five computing platforms evaluated. This parallelization of ML has enabled tomosynthesis to become a viable technology in the breast screening clinic, reducing reconstruction time from 3 hours on a PentiumIVworkstation to 6 minutes on a 32-node PentiumIV cluster.