Abstract
Objectives: The objective of this study is to investigate the clinical potential of generating artificial bone scintigraphy (aBS), defined here as a deep learning-generated bone scintigraphy image that simulates delayed-phase bone scintigraphy (dBS) characteristics, from early-phase bone scintigraphy (eBS) obtained with a short waiting time using an unpaired image-to-image translation method in patients with breast cancer (BC). Methods: In this single-center prospective study involving 245 patients with BC (195 for training and 50 for testing), eBS and dBS were performed. Using the contrastive unpaired translation (CUT) model, we trained with anterior and posterior images of the eBS and dBS from the training group. We then generated aBS images targeting dBS by inputting eBS from the test group for both anterior and posterior views. We conducted quantitative, qualitative, and visual assessments to evaluate aBS. Results: The points of the anterior and posterior images of aBS on the qualitative four-point and five-point rating scales were significantly higher than those of eBS (p < 0.0001). Three nuclear medicine physicians performed visual assessments, demonstrating consistent findings on the presence of bone metastases in both aBS and dBS. Their visual evaluations indicated that the bone-to-soft tissue contrast in aBS was superior to that in eBS. The quantitative metrics of aBS were superior to those of eBS. However, aBS was inferior to the targeted dBS in terms of qualitative and visual assessments. Conclusions: The aBS generated through CUT was superior to eBS in quantitative, qualitative, and visual assessments and preserved lesion-related information comparable to dBS. Although these findings do not support replacement of dBS for definitive diagnosis, they support the feasibility of aBS as an assistive delayed-phase-like image generation approach from earlier-acquired bone scintigraphy.