Approximately one third of cancer patients are diagnosed with vertebral metastasis and associated bone quality impairment. Treatments for metastatic disease have proven beneficial to overall prognosis, yet their effects on bone quality are not fully characterized. This research utilizes image analysis, physical testing, and generative deep learning (DL) approaches to characterize bone quality and treatment-associated changes in a preclinical rat model of osteolytic vertebral metastasis. The specific aims are to enhance barium sulfate (BaSO4) labelled microdamage visualization in μCT images of vertebrae, determine tissue level material and mechanical properties of healthy and osteolytic vertebrae with and without treatment, and develop generative DL models predicting vertebral fracture.

μCT image acquisition parameters were iteratively adjusted to enhance BaSO4 contrast and spatial resolution. Enhanced μCT image acquisition parameters yielded BaSO4 labels with high probability of pixelwise spatial correlation (83% within 20 μm) to microdamage evaluated in gold-standard backscatter electron (BSE) images. This work provides a high-resolution protocol for 3D μCT analysis of microdamage accumulation in vertebral bone.

Microdamage accumulation, load-to-failure, and microstructural parameters were measured in healthy and osteolytic rat vertebrae with and without treatment (stereotactic body radiotherapy (SBRT), docetaxel, zoledronic acid). Microdamage was increased in metastatic vertebrae and was reduced by all treatments. Load-to-failure was decreased in untreated and SBRT tumor-injected rats compared to healthy controls. Strong correlations were found between microstructural parameters, load-to-failure and microdamage accumulation.

A DL approach was developed to produce sample-specific synthetic μCT images of rat vertebrae undergoing destructive testing. The cGAN generated realistic 3D µCT images of rat vertebrae through the fracture process. This motivated the development of a hybrid modeling approach using a dual encoder cGAN to augment input μCT images with physics models, including axial rigidity and linear elastic strain. The axial rigidity model significantly improved fracture prediction precision compared to the baseline model, with the F1-score trending to significance. This work has provided recommendations for future DL approaches using improved loss functions for identifying generated fractures.

This research has contributed new imaging, physical testing, and DL methods for assessing bone quality in the metastatic spine and has quantified the impact of disease and clinically relevant cancer treatments. The outcomes of this work may better inform treatment planning to prevent bone quality degradation in patients with vertebral metastatic disease.