Software-based de-filtering restores quantitative accuracy in Clarity2D-enhanced whole-body bone scintigraphy.
Authors
Affiliations (5)
Affiliations (5)
- Department of Diagnostic and Interventional Radiology, University of Tsukuba Hospital, 1-1-1 Tennodai, Tsukuba, 305-8575, Ibaraki, Japan.
- Department of Radiology, Institute of Medicine, University of Tsukuba, 1-1-1 Tennodai, Tsukuba, 305-8575, Ibaraki, Japan.
- Department of Radiology Technology, University of Tsukuba Hospital, 1-1-1 Tennodai, Tsukuba, 305-8575, Ibaraki, Japan.
- Department of Radiology, Institute of Medicine, University of Tsukuba, 1-1-1 Tennodai, Tsukuba, 305-8575, Ibaraki, Japan. [email protected].
- Center for Cyber Medicine Research, University of Tsukuba, 1-1-1 Tennodai, Tsukuba, 305-8575, Ibaraki, Japan. [email protected].
Abstract
To determine whether software-based de-filtering can restore the quantitative accuracy of the bone scan index (BSI) and the number of hot spots (HSn) in whole-body scintigraphy images degraded by the Clarity2D noise-reduction filter. In this IRB-approved retrospective study, 101 adults (mean age ± SD: 67 ± 13 years) who underwent <sup>99m</sup>Tc-HMDP whole-body scintigraphy on a cadmium-zinc-telluride (CZT) SPECT/CT system were analyzed. For each patient, three planar datasets were obtained: (i) unfiltered images, (ii) 40%-blend Clarity2D-filtered images, and (iii) software de-filtered images reconstructed with a deep learning-based inverse filter in VSBONE BSI v3.0. Quantitative indices (BSI and HSn) and lesion masks were automatically extracted. Agreement with the unfiltered reference was evaluated using Pearson correlation, Bland-Altman analysis (bias ± 95% limits), Dice coefficient, and the Hausdorff distance (p < 0.05). Additionally, lesion detection accuracy was quantified using intersection over union (IoU)-based matching to calculate precision, recall, and F1-score. Clarity2D filtering significantly impaired quantitative concordance (BSI r = 0.23, bias = - 1.55 [-6.20 to 3.10]; HSn r = 0.23, bias = - 14.4 lesions). In contrast, de-filtering restored concordance (BSI r = 0.99, bias = - 0.04 [-0.26 to 0.17]; HSn r = 0.98, bias = - 0.04 lesions) and improved spatial overlap (Dice 0.40 to 0.82) while reducing the median Hausdorff distance from 103 pixels (IQR 85-188) to 39 pixels (IQR 1-40) (all p < 0.001). The de-filtered method demonstrated superior lesion detection accuracy compared to Clarity2D (Precision: 0.77 ± 0.37 vs. 0.19 ± 0.29, Recall: 0.81 ± 0.37 vs. 0.43 ± 0.44, F1-score: 0.78 ± 0.36 vs. 0.23 ± 0.30). Furthermore, de-filtering achieved high inter-case stability (median F1-score: 1.0), whereas Clarity2D showed substantial variability (median F1-score: 0.057). The proposed de-filtering algorithm reliably reverses Clarity2D-induced distortions, enabling accurate BSI and HSn measurements and robust lesion detection without additional radiation or acquisition time. This technique has the potential to broaden the clinical adoption of noise-reduction filters while preserving the integrity of downstream quantitative analyses.