Discrete color-based Euclidean-invariant signatures for feature tracking in a DIET breast cancer screening system

Abstract

A Digital Image-based Elasto-Tomography (DIET) system for breast cancer screening has been proposed in which the elastic properties of breast tissue are recovered by solving an inverse problem on the surface motion of a breast under low frequency (50-100 Hz) mechanical actuation. The proposed means for capturing the surface motion of the breast in 3D is to use a stroboscope to capture images from multiple digital cameras at preselected phase angles. Photogrammetric techniques are then used to reconstruct matched point features in 3D. Since human skin lacks high contrast visual features, it is necessary to introduce artificial fiducials which can be easily extracted from digital images. The chosen fiducials are points in three different colours in differing proportions randomly applied to the skin surface. A three-dimensional signature which is invariant to locally Euclidean transformations between images is defined on the points of the lowest proportion colour. The approximate local Euclidean invariance between adjacent frames enables these points to be matched using this signature. The remaining points are matched by interpolating the transformation of the matched points. This algorithm has significant performance gains over conventional gradient-based tracking algorithms because it utilises the intrinsic problem geometry. Successful results are presented for simulated image sequences and for images of a mechanically actuated viscoelastic gel phantom with tracking errors within 3 pixels. The errors in the phantom sequence correspond to less than 0.3 mm error in space, which is more than sufficient accuracy for the DIET system.