The inversion procedure presented in this paper is based on the statistical regression of the inverse map between the spaces of ECT scan data, and of crack parameters. The mapping is realized by a combination between a statistical data processing step, i.e., a principal component transformation of the scan data, and an incremental resolution neural network training. Starting from the necessities of improving the detrimental conditioning of the regression and of providing the inversion approach with enhanced potential for automation, a novel “shifting aperture” mapping concept and a data fusion technique are proposed. Supplementing the primary mapping algorithm with these latter processing steps allows one to avoid the usual anomalous-region focusing approach and improves the inversion capabilities by allowing adynamic reconstruction of the object's profile. Unconnected and multiply connected crack shapes are well estimated, that so far eluded most other inversion methods. For this primary validation of the completed algorithm, only synthetic B-scan data are used, which are collected by an optimized, high performance sensor on the interior of a metal tube.