Eardrums are automatically classified based on a feature set from a three-dimensional image of the eardrum, such as might be derived from a plenoptic image captured by a light field otoscope. In one aspect, a grid is overlaid onto the three-dimensional image of the eardrum. The grid partitions the three-dimensional image into cells. One or more descriptors are calculated for each of the cells, and the feature set includes the calculated descriptors. Examples of descriptors include various quantities relating to the depth and/or curvature of the eardrum. In another aspect, isocontour lines (of constant depth) are calculated for the three-dimensional image of the eardrum. One or more descriptors are calculated for the isocontour lines, and the feature set includes the calculated descriptors. Examples of descriptors include various quantities characterizing the isocontour lines.