Real-time ultrasound images from the Longissimus dorsi muscle across 11th to 13th ribs of 720 live bulls and steers were acquired over the period of four years. The actual intramuscular percentage of fat (IFAT) was determined using an n-hexane extraction with mean of 4.98%, standard deviation of 2.12%, and range from 1.10% to 14.68%. Image-processing techniques were used to calculate parameters to quantify the image texture patterns. The parameters which showed good correlations with the actual IFAT were used to develop a statistical linear regression model. The accuracy of prediction was very good for the actual IFAT less than or equal to eight (low IFAT group), with root mean square error (RMSE) around 1.0%. However, the model was much less accurate for prediction of IFAT values more than eight (high IFAT group), with RMSE more than 1.5%. One reason for this could be the limited ability of the ultrasound technique to resolve differences in high-IFAT muscles in terms of image texture patterns. Also, this group contained fewer than 10% of the images collected, which may be an inadequate sample. Overall accuracy of prediction was improved by developing different regression models for the low-IFAT and high-IFAT groups. Statistical pattern recognition and classification techniques were applied to “pre-classify” the images into low- or high-IFAT groups before being subjected to regression prediction models. The techniques applied included cluster analysis, discriminant analysis, and classification and regression tree (CART). The classification tree provided the best results with overall classification accuracy around 90% for low- and high-IFAT groups of images. In conclusion, overall accuracy of predicting the IFAT from ultrasound image parameters and regression models can be improved by first isolating the high- IFAT group from low-IFAT group using statistical classification methods.