BMD-GAN: Bone Mineral Density Estimation Using X-Ray Image Decomposition into Projections of Bone-Segmented Quantitative Computed Tomography Using Hierarchical Learning

Assessing the utility of osteoporosis self-assessment tool for Asians in patients undergoing hip surgery

“…While other screening tools such as SCORE [ 4 ], ORAI [ 5 ], ABONE [ 6 ], and COSA [ 7 ] may partially serve as the solution, artificial intelligence (AI) technology has recently been applied for evaluating X-rays to screen for osteoporosis. In fact, some studies using AI have shown remarkable accuracy with r > 0.9 in measuring BMD and AUC > 0.9 in detecting osteoporosis from X-ray images [ [23] , [24] , [25] ]. However, concerns regarding the robustness and accessibility of AI-developed models have emerged.…”

Bone mineral density estimation from a plain X-ray image by learning decomposition into projections of bone-segmented computed tomography

“…Recent studies have focused on the use of deep learning to estimate BMD or diagnose osteoporosis from plain X-ray images (Hsieh et al, 2021;Yamamoto et al, 2020;Jang et al, 2021;Ho et al, 2021;Gu et al, 2022) (Wang et al, 2023), which are more widespread modalities than DXA and QCT. Most existing methods directly regress BMD or classify osteoporosis from regions of interest (ROI) of X-ray images, some of which achieved high correlations with DXA-measured BMD (Hsieh et al, 2021;Wang et al, 2023).…”

“…• Extensive validation with real clinical datasets involving variations in patient poses, image compression ratios, and CT calibration. (Gu et al, 2022) and then learns mapping to BMD by regressing from the learned representation. (Right) The visualized procedure of shaping representation and model attention shifting with learning stages.…”

“…(Right) The visualized procedure of shaping representation and model attention shifting with learning stages. The representations were visualized after being compressed to one-channel images using principal component analysis; the model attention maps were visualized using FullGrad (Srinivas and Fleuret, 2019) This paper is built upon our conference paper (Gu et al, 2022) with extensions into several aspects. From the methodological aspect, we explicitly targeted QCT-BMD in addition to DXA-BMD.…”

“…We validated the effectiveness of the proposed decomposition learning by regressing BMD from the learned representation with a transformer-based regression network. For the dataset aspect, we included more patients who had undergone X-ray scanning and QCT scanning and had taken BMD measurements, resulting in 305 patients (increased from 200 patients) in Gu et al (2022). We included more Xray images imaged with supine, abduction, and adduction poses in addition to the standing pose.…”