Use of articulated registration for response assessment of individual metastatic bone lesions.

Accurate skeleton registration is necessary to match corresponding metastatic bone lesions for response assessment over multiple scans. In articulated registration (ART), whole-body skeletons are registered by auto-segmenting individual bones, then rigidly aligning them. Performance and robustness of the ART in lesion matching were evaluated and compared to other commonly used registration techniques. Sixteen prostate cancer patients were treated either with molecular targeted therapy or chemotherapy. Ten out of the 16 patients underwent the double baseline whole-body [F-18]NaF PET/CT scans for test-retest (TRT) evaluation. Twelve of the 16 patients underwent pre- and mid-treatment [F-18]NaF PET/CT scans. Skeletons at different time points were registered using ART, rigid, and deformable (DR) registration algorithms. The corresponding lesions were contoured and identified on successive PET images based on including the voxels with the standardized uptake value over 15. Each algorithm was evaluated for its ability to accurately align corresponding lesions via skeleton registration. A lesion matching score (MS) was measured for each lesion, which quantified the per cent overlap between the lesion's two corresponding contours. Three separate sensitivity studies were conducted to investigate the robustness of ART in matching: sensitivity of lesion matching to various contouring threshold levels, effects of imperfections in the bone auto-segmentation and sensitivity of mis-registration. The performance of ART (MS = 82% for both datasets, p ≪ 0.001) in lesion matching was significantly better than rigid (MS(TRT)=53%, MS(Response)= 46%) and DR (MS(TRT)=46%, MS(Response)=45%) algorithms. Neither varying threshold levels for lesion contouring nor imperfect bone segmentation had significant (p~0.10) impact on the ART matching performance as the MS remained unchanged. Despite the mis-registration reduced MS for ART, as low as 67% (p ≪ 0.001), the performance remained to be superior to the rigid and DR algorithms. ART is not only robust to contouring threshold levels for bone lesions, but also outperforms rigid and DR algorithms in lesion matching. ART therefore enables the study of TRT variability and treatment assessment of individual bone lesions.