It has been shown the application of Mean-Shift Clustering (MSC) to fMRI analysis can increase detection sensitivity in low contrast to noise (CNR) situations (Ai et al. 2014). The feature space previously used with MSC was based purely on spatial features of the data. In this study, MSC was utilized with a feature space containing both temporal and spatial features to further increase its detection power. The proposed MSC was applied to simulated and real fMRI data to examine its sensitivity and specificity and compared to the commonly used cluster analysis (CA). If successful, the proposed technique can improve detection in techniques that inherently have low CNR, such as non-proton fMRI or non-BOLD fMRI.The new feature space used with MSC for this study is the estimated mean Z values surrounding a voxel and the signal power at the base frequency of the experimental design. The mean Z values surrounding a voxel were used to take into consideration neighboring effects while the power at base frequency takes into consideration temporal characteristics. The MSC technique was evaluated with (MSC+CA) and without additional CA applied. The techniques will be applied to a statistical parametric image generated using cross correlation analysis (CCA). A cluster size threshold of 4 voxels was used for the application CA when applied. The new feature space was tested on a simulated fMRI data sets (100 images, a 128x128 matrix) while simulating a block design with 2.5 off/on cycles and an event related design with a stimulation every 4 images. Receiver Operating Characteristic (ROC) curves were generated for evaluating the efficacy of the technique at various CNRs (0.20, 0.40, 0.60, 0.80). The same MSC technique was also tested on real BOLD fMRI data acquired on a Siemens 3T Trio scanner (Siemens Medical Solutions, Erlangen, Germany) using electrical stimulations delivered to the right median nerve with a Grass S48 stimulator (Grass Technologies, USA). The stimulation intensity was adjusted to 15V above the minimum required to obtain a thumb twitch. A block design with four and a half off/on cycles was used with a randomized inter-stimulation interval (ISI) of 1.5-2.5 seconds with a length of 180 images. The volunteers were asked to stay still and awake for the duration of the scan with no task actively performed during the scan. The images were processed using Analysis of Functional NeuroImages (AFNI) and custom written Matlab code.The results indicate that the proposed MSC technique can improve activation detection over CA based on the ROC curves (Figure 1) generated from simulated data, with MSC+CA improving detections even more. Using a kernel size of 0.30 in both block and event related designs, improvements can be seen over CA with the MSC techniques in the lower CNR situations (0.20 and 0.40). Activations can be seen when the same method is utilized on real BOLD fMRI data (Figure 2) with detected activations being apparently larger when compared to CA at the same significance level.The results in the ROC curves with the simulated data using MSC show an improvement in detection over CA in low CNR situations, even more so if using MSC+CA. Since activation detection at high CNR scenarios typically does not present an issue but does at low CNR scenarios, utilizing MSC with temporal characteristics can improve activation detection in such situations. Analysis techniques that target very low CNRs could be particularly useful for imaging studies utilizing methods with inherently low CNRs, non-proton imaging or non-BOLD fMRI for instance. MSC may also be able to shorten acquisition time required for currently used BOLD fMRI methods by reducing the number of repetitions required. Further optimization studies need to be performed on this technique, on kernel sizes, different clustering techniques to use in conjunction, or different feature spaces for example.