We study the accuracy of reconstruction of a family of functions fϵ(x), x ∈ R 2 , ϵ → 0, from their discrete Radon transform data sampled with step size O(ϵ). For each ϵ > 0 sufficiently small, the function fϵ has a jump across a rough boundary Sϵ, which is modeled by an O(ϵ)-size perturbation of a smooth boundary S. The function H0, which describes the perturbation, is assumed to be of bounded variation. Let f rec ϵ denote the reconstruction, which is computed by interpolating discrete data and substituting it into a continuous inversion formula. We prove that (f rec ϵ − Kϵ * fϵ)(x0 + ϵx) = O(ϵ 1/2 ln(1/ϵ)), where x0 ∈ S and Kϵ is an easily computable kernel.