To illustrate with a more specific example, FIG. 3A shows an exemplary H&E image containing various types of regions, such as the necrosis region, the cancer (tumor) region, the lymphocytes region, and the background region. FIG. 3B shows the same image being overlaid with boundaries of exemplary superpixels generated for the image by superpixel generator 110. In this example, the number of generated superpixels is N=372. Assuming now that for each superpixel, the following image features are generated for each of the three color channels R, G, and B: a histogram of the intensity, a histogram of the gradient magnitude, and a histogram of the gradient direction. If each histogram includes 10 bins, for example, then the total number of features extracted for each superpixels is M=10×3×3=90. That is, each superpixel can be characterized by a ninety-dimensional feature vector. Clustering engine 112 can then be configured to cluster (e.g., using k-means) the feature vectors into K=10 clusters, resulting in ten ninety-dimensional centroid vectors representing the ten clusters. In this example, a compression ratio of 372×90/(372+10×90+10×9/2)=25.42 can be achieved if centroid vectors are stored, and an even higher compression ratio of 372×90/(372+10×9/2)=80.28 can be achieved if centroid vectors are not stored.