It is projected that there will be 1.9 million new cancer diagnoses and over 600,000 cancer deaths in 2022 in the United States alone.¹ Spine is the most common site for bone metastasis; up to 70% of cancer patients could experience spinal metastasis (SM) and about 20% of them become symptomatic, causing considerable pain and morbidity.² Most common regions for SM are dorsal (45%), lumbar (17%), cervical (14%), and dorsolumbar (10%).³ Vertebral fractures (VFs) are the common complications of SM, with an estimation of 30% of patients developing VFs that mostly require surgery.⁴ Tumor lesions and some treatments for cancer can lead to loss of bone tissue, thus increasing the risk of VFs.⁵ Hence, it is imperative to predict the VF risk in SM patients, so that well-informed interventions can be designed.⁵ Factors influencing VFs include both the macro- and micro-structure of the vertebra, especially of the trabecula.⁵ In-vitro methods are used to study the biomechanical forces that alter the vertebral shape, but they cannot accurately measure in-vivo stresses that damage its microstructure.⁵ The development of computational modeling has made it a compelling tool for measuring these in-vivo stressors.⁵ A recent study by Ahmadian et al was to assess the feasibility of the artificial intelligence (AI)-assisted framework to create ‘digital twins’ of the human vertebra, termed ReconGAN, to predict the VF risk in both osteolytic and osteoblastic metastatic tumors.⁵

A digital twin is an in silico version of a physical entity, and when applied to orthopedics, it has the potential to simulate disease progression and therapy.^6,7 However, most studies did not take into account the trabecular microstructure, which is highly indicative of fracture risk.⁵ It is essential to generate a realistic finite element (FE) model of this microstructure for accurate VF prediction.⁵ Micro-computed tomography (CT) scans provide submicron level 3D bone microstructure that the diagnostic CT scans cannot.⁵ However, stringent sample preparation requirements for micro-CT scans make it impossible to be used for in vivo studies.⁵ Thus, most current FE studies lack the microstructure perspective on the biomechanics of the human spine.⁵

Currently, algorithms used for virtual reconstruction are too simplistic to represent the complex microstructure and the materials of the trabecula.⁵ Generative models and deep learning (DL) algorithms, like convolutional neural networks, can instead to be used to create realistic microstructures using imaging data.⁵ However, these algorithms require large and diverse training data which are yet available for trabecular microstructure.⁵ This challenge can be circumvented by a class of DL called generative adversarial networks (GANs).⁵ Deep Convolutional GANs (DCGANs) have been used to create 2D images of liver lesions which were used as a training set for other DL models for lesion classification.⁸

Ahmadian et al developed an AI-assisted framework called ReconGAN, which employs 3D DCGANs trained on quantitative micro-CT images obtained from cadaveric vertebra samples to create models of the trabecular bone.⁵ Notably, in view of scarcity of training samples, they used a data augmentation method using DCGAN for training and used only 1 cadaveric vertebra sample.⁵ They created a whole vertebra model by engraving this structure within the cortical shell extracted from the patient's CT scan.⁵ There was an abrupt transition between the trabecular and the cortical region, which gave false excessive stress concentrations during the simulation, leading to underprediction of vertebra strength,

and was then removed by an FE-based shape optimization.⁵ Finally, this model was converted to a high-fidelity FE-model, and its fracture response to metastatic tumors was predicted.⁵

It was found that under even compression, the damage started from the weaker areas of the trabecula and propagated to the cortical shell forming a macro-crack.⁵ The energy absorbed by the trabecula during this phenomenon caused a ductile failure response.⁵ Under high flexion loads, amplified compressive stress in the anterior cortical region led to brittle failure response and a significant drop in the vertebra height in the anterior cortical shell, which was also confirmed by clinical observations.⁵ Additionally, incorporating an identically-shaped lytic and a blastic tumor model inside the trabecula caused a 28% and 13% reduction in the vertebra strength, respectively.⁵

This study highlighted the crucial role of modeling trabecular microstructure and macroscopic shape using ReconGAN in predicting VFs under uniform compression or high flexion loads.⁵ This feasibility study demonstrated that ReconGAN was a reliable tool which could be used in creating patient-specific models for evaluating VF risk and designing appropriate interventions after further comprehensive studies.⁵