The backscatter coefficient (BSC) describes the scattering properties of a medium and can be used to characterize tissue, such as fatty liver. To calculate the BSC, a calibration spectrum is needed, which can be acquired using a reference phantom method before or after the clinical scanning procedure is complete. This requires yet another scanning procedure that can disrupt the busy clinical flow. Therefore, we have explored the use of a convolutional neural network (CNN) as a means of eliminating an external reference step while still capable of classifying liver disease in a rabbit model of fatty liver. Sixty New Zealand white rabbits were separated into five cohorts with each cohort maintained on a special high fat diet to induce different degrees of fatty liver. One week before scanning, rabbits were placed on normal chow. Rabbit livers were scanned in vivo using an L9-4/38 linear array connected to a SonixOne scanner. Raw RF data were collected from the scans and used for quantitative ultrasound (QUS) analysis. Immediately after scanning, the rabbit liver was extracted and the percent lipids in the liver were estimated using the Folch assay. The rabbit livers were classified into two classes: high fat (lipid percentage >= 5%) and low fat (lipid percentage <5%). The 5% threshold was equal to the median of lipid percentages of all the rabbits. Livers were classified using traditional QUS approaches and compared with a CNN approach where no reference was utilized, and the raw RF signals were used as the inputs. The attenuation and the BSC were estimated from the RF using a reference phantom technique. The attenuation slope, attenuation midband-fit, the effective scatterer diameter, and the effective acoustic concentration were extracted from the attenuation and the BSC and used as features to train a kernel support vector machine (SVM) for the task of liver lipid classification. A CNN was designed to simultaneously extract the features from the raw RF and perform liver lipid classification without using a reference phantom. Six-fold cross validation was performed to quantify the accuracy of the SVM classifier using the QUS parameters and the CNN classifier using the RF. The average training and testing accuracies across six folds using the QUS approach was 68.94% and 59.12%, respectively. The average training and test accuracy using the CNN approach were 81.03% and 73.81% for training and testing, respectively. The results demonstrate that the CNN can be used to classify fatty liver without the need for an external reference scan, i.e., reference-free QUS.