The retina is a layered structure, and some diseases can be clinically predicted and diagnosed based on the change in the thickness of the retinal layer. To segment the different layers of the retina quickly and accurately, this study proposes a random forest algorithm based on principal component analysis (PCA). The algorithm uses PCA to resample the normalized features collected from the retinal images and retains the feature information dimensions with significant weight, thereby eliminating the relevance between the different feature dimensions and information redundancy. After PCA, the number of features can be reduced obviously, but still retains 99% information. Random forests algorithm applies the features to learn and predict the location of retinal layer boundaries. We extract each pixels values of retinal boundaries, producing an accurate probability map for each boundary. Experimental results show that when the total number of feature dimensions decreased from 29 to 18, the training speed of the model increased by 23.20%. By contrast, when the number of feature dimensions was 14, the training speed increased by 42.38%. However, the effect on image segmentation accuracy was not obvious. Thus, it is found that this method effectively improves the efficiency of the algorithm.