Fig. 2: Occlusion Sensitivity Mapping (OSM) highlights image areas with high importance for correct class predictions, enabling output interpretation.

OSM iteratively blocks image areas from being evaluated by the deep learning network. If an image area is highly important for classification, the network’s performance will thus drop substantially in the given iteration. Importance is thus standardized between 0 and 1, where values between 0.5 and 1.0 denote high importance, i.e., a sharp drop in classification performance if that image area is blocked (red, see color bar). A standard field of view of bone marrow smears from MDS patients is shown in (A, C, and E). The corresponding OSM is displayed in (B, D, and F), respectively. First, in a proof-of-concept fashion, the networks focus on cells and specifically on nuclei. Background, noise, or smudge are not considered important for classification. Second, high importance is given to erythropoietic and granulopoietic cells as well as megakaryocytes.