Figure 1.

(Row 1) The robust multi-level features used to help detect the very small tumor region. From left to right, (a) original image containing the very small tumor region, (b) zoomed-in view of the tumor region, (c) the ground-truth, (d) the result with multi-level distance features, and (e) the result without those features. (Row 2) Meaningful bone tumor matrices of knee bone tumor classification shown in (f–h). It is proven to be a highly predictive feature of bone tumor classification in [5]. This explains why global and patch-based approaches should be applied to distinguish between benign-tumor and malignant-tumor regions.