Abstract
Quantitative assessment of metastatic involvement of the bony spine is important for assessing disease progression and treatment response. Quantification of metastatic involvement is challenging as tumours may appear as osteolytic (bone resorbing), osteoblastic (bone forming) or mixed. This investigation aimed to develop an automated method to accurately segment osteoblastic lesions in a animal model of metastatically involved vertebrae, imaged with micro computed tomography (μCT).
Radiomics seeks to apply standardized features extracted from medical images for the purpose of decision-support as well as diagnosis and treatment planning. Here we investigate the application of radiomic-based features for the delineation of osteoblastic vertebral metastases. Osteoblastic lesions affect bone deposition and bone quality, resulting in a change in the texture of bony material physically seen through μCT imaging. We hypothesize that radiomics based features will be sensitive to changes in osteoblastic lesion bone texture and that these changes will be useful for automating segmentation.
Osteoblastic metastases were generated via intracardiac injection of human ZR-75-1 breast cancer cells into a preclinical athymic rat model (n=3). Four months post inoculation, ex-vivo μCT images (µCT100, Scanco) were acquired of each rodent spine focused on the metastatically involved third lumbar vertebra (L3) at 7µm/voxel and resampled to 34µm/voxel.
The trabecular bone within each vertebra was isolated using an atlas and level-set based segmentation approach previously developed by our group. Pyradiomics, an open source Radiomics library written in python, was used to calculate 3D image features at each voxel location within the vertebral bone. Thresholding of each radiomic feature map was used to isolate the osteoblastic lesions.
The utility of radiomic feature-based segmentation of osteoblastic bone tissue was evaluated on randomly selected 2D sagittal and axial slices of the μCT volume. Feature segmentations were compared to ground truth osteoblastic lesion segmentations by calculating the Dice Similarity Coefficient (DSC). Manually defined ground truth osteoblastic tumor segmentations on the μCT slices were informed by histological confirmation of the lesions.
The radiomic based features that best segmented osteoblastic tissue while optimizing computational time were derived from the Neighbouring Gray Tone Difference Matrix (NGTDM). Measures of coarseness yielded the best agreement with the manual segmentations (DSC=707%) followed by contrast, strength and complexity (DSC=6513%, 5428%, and 4826%, respectively).
This pilot study using a radiomic based approach demonstrates the utility of the NGTDM features for segmentation of vertebral osteoblastic lesions. This investigation looked at the utility of isolated features to segment osteoblastic lesions and found modest performance in isolation. In future work we will explore combining these features using machine learning based classifiers (i.e. decision forests, support vector machines, etc.) to improve segmentation performance.
