4-D lesion detection using expectation-maximization and hidden Markov model

We explored the use of spatial and temporal information to automatically detect lesions in 4D medical image data (3D space + 1D time). The 3D method, expectation-maximization (E-M), was used to estimate the probability distributions of various tissue classes in the image. Evolution of the lesion over time is assumed to follow a hidden Markov model (HMM), where the state of the system is expressed as lesion or non-lesion independently for each voxel in the image. Synthetic images based on a Gaussian mixture model were used to simulate a 4D image data set with exponential lesion growth. A comparison was made between use of the E-M algorithm alone, run independently on each image in the time series, and the E-M plus HMM temporal approach. The combination of techniques showed improvement in sensitivity and specificity of lesion detection.