Directional edge registration for temporal chest image subtraction

We used a directional filtering technique to accurately register the ribs on temporal chest radiographs. Rib registration was the primary technical objective. In order to accurately extract the ribs, we developed a directional edge function that used rotating kernels through a dynamic search technique to extract rib edges of varying contrast. A directionally-weighted operator with a phase-contrast feature is embedded in the kernel design, which allows us to separate the upper and lower bounds of structures in the vertical direction. Broken rib edges were reconnected using a reasoning algorithm by analyzing the rib position and curve. The lung and heart boundaries were extracted using an inverted umbrella filter. Control points were designated at turning points and distributed on smooth edge segments on the temporal images. A thin-plate spline technique took the control points and performed the final registration prior to the temporal image subtraction. We found that normal chest structures (i.e. ribs, heart and normal lung structures) were greatly reduced in the subtraction images. In addition, the contrast of cancer spots was significantly increased.