MRI grayscale image data pattern, in reality, is not always symmetrical. Sometimes, it has a skewed pattern, leptokurtic, mesokurtic, platykurtic and even fat-tail in distribution. The Gaussian approach is not always able to explain this kind of data pattern. Therefore, other approaches to employ other distributions are used to overcome this problem. This study tries to compare the Gaussian, Student's t, and Laplacian mixture model in the modeling of MRI brain tumor image segmentation. In addition, we used the Markov Random Field as the prior to the spatial dependence. The cluster validation is done by calculating the Silhouette Index (SI) and Misclassification Ratio (MCR). The results demonstrate that for the skewed and leptokurtic data pattern, the Laplacian mixture model shows the best representation, while the Student's t mixture model has a great performance in fat-tail data pattern and also more robust against noise.