Existing public datasets for osteoporosis screening remain a significant bottleneck. (1) Lack anatomical imaging; (2) Lack paired conventional X-ray or CT images; (3) Lack standardized BMD reference values. These disparities highlight the urgent need for more comprehensive, standardized, and multimodal public datasets to advance osteoporosis research and improve diagnostic practices.

Here are some examples of X-ray and CT images in the LUMOS dataset. CT images of an osteoporotic case reveal a more porous trabecular structures and a thinner cortical layer in comparison to both a normal case and an osteopenia case.

We test various osteoporosis screening methods on the LUMOS dataset, with accuracy (ACC) and F1 score (F1) metrics, and the results are shown in the below Table. X-rays, accessible even in resource-limited areas, can be analyzed by these machine learning algorithms to identify features like cortical bone thinning, trabecular pattern alterations, and fractures, offering a cost-effective initial screening method. CT scans, with their high-resolution cross-sectional views, allow for the development of sophisticated models that detect subtle microarchitectural changes not easily visible on X-rays, facilitating more accurate and nuanced diagnosis. Multimodal models, combining X-ray, CT, leverage the unique strengths of each data source to capture complex relationships between imaging features and patient factors, achieving higher accuracy in diagnosing and prognosticating osteoporosis compared to single-modality methods.

We conduct BMD prediction experiments on multiple methods using LUMOS' multimodal data, and measure prediction accuracy through metrics such as Root Mean Square Error (RMSE) and Pearson Correlation Coefficient (PCC). The results are shown in the above table. By analyzing visual cues such as bone radiolucency and trabecular pattern density in lumbar X-rays, models trained on this dataset can estimate BMD, offering a viable option for initial assessments. The CT modality, with its high-resolution capabilities, enables in-depth analysis of bone microarchitecture, allowing for more precise BMD prediction. The dataset's multimodal nature truly shines when it comes to BMD prediction, as multimodal models can integrate the quick initial assessments from X-rays, the detailed structural analysis from CTs, the reference values from DXA, and the patient-specific factors from metadata. This comprehensive integration overcomes the limitations of individual modalities, delivering highly accurate and reliable BMD predictions that enhance osteoporosis management and improve patient outcomes.

We use the point-prompt and box-prompt interface of Segment Anything Model (SAM) and demonstrate robust automated segmentation results on LUMOS images, as illustrated in the above figure. Although these approaches may not match the accuracy of fully-supervised methods with high-fidelity ground truth, they validate the LUMOS dataset’s utility for segmentation research and highlight its potential to drive innovation in medical image segmentation.

In regions with scarce medical resources, synthesizing CT images from X-rays holds significant clinical value. The previous study like X2CTGAN has attempted to use GANs to generate CT from two orthogonal X-rays. However, the scarcity of paired X-ray-CT datasets in medical research has forced reliance on synthetic X-rays created via CT-derived DRR (Digital Reconstructed Radiography) projection for training and validation. The LUMOS dataset addresses this critical gap by providing paired CT and orthogonal X-ray images, offering a rare real-world dataset for X-ray to CT image synthesis tasks. What's more, inspired by recent advancements in neural radiance fields (NeRRFs) for 3D image generation from a single natural image, LUMOS further enables the exploration of NeRF-based methods to generate 3D CT-like volumetric data from single-view X-rays, overcoming the need for multi-angle acquisitions. LUMOS' aligned, multimodal imaging of the lumbar spine provides the essential foundation for developing and validating these innovative techniques, driving forward the translation of AI-driven imaging solutions to resource-constrained environments.

The LUMOS dataset encompasses a wide variety of images, with differences in magnetic field strengths, number of slices, slice thickness, and scanner manufacturers. This diversity makes it an ideal resource for developing domain generalization and image standardization techniques. Image standardization aims to transform images from different sources into a consistent format, reducing variability and improving the performance of AI models across different datasets and imaging devices. Researchers can explore methods to normalize image intensity, align anatomical structures, and standardize image resolutions using the LUMOS dataset. These can lead to the development of more effective image analysis techniques and potentially improve the diagnosis and treatment of various diseases, not just limited to osteoporosis.