Transferable HMM Trained Matrices for Accelerating Statistical Segmentation Time

Segmentation problem has been solved in the last decade, many techniques have been implemented to discover this issue. Many problems arise during the segmentation process including the acceptable error rate, low quality assessment, and time complexity. A variety of acceleration techniques have been applied to speed up the segmentation time and achieve a segmented result in real time. GPU and parallel processing using hardware have been employed efficiently here, but still limited in 3D images segmentation. Hidden Markov Model (HMM) is one of the best statistical segmentation techniques that played a significant rule recently. The problem associated with HMM is the time complexity due to the training steps. This issue has been resolved using different accelerator but still not efficient with 3D volumes. In this research, we propose a methodology for transferring the trained matrices of HMM from image to another skipping the training time for the rest of the 3D volume. One HMM train is generated and generalized to the whole volume. An accurate segmentation results have been achieved in less processing time. And fixed class belongings for the pixels have been achieved without any class membership variations. Which will increase the possibility of segmenting medical images using HMMs on GPUs instead of CPUs.