5 Fractured Vertebrae

[Orencio Suhag]

Thankyou for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

Radiomics is a multi-step process of converting medical images into meaningful and mineable data1,2. In the hand-crafted radiomics pipeline, the process includes segmentation, feature extraction, feature selection, and construction of diagnostic, prognostic or predictive models. Radiomics has shown promising results in oncologic imaging as a tool to reflect the tissue heterogeneity and its application to other medical fields, including spine imaging, has been growing1,2.

Differentiation of benign and malignant compression fracture is a frequently encountered problem in clinical practice. Accurate diagnosis is important with a considerable difference in management and prognosis. In particular, it is increasingly important but challenging to differentiate benign and malignant fractures in elderly population with both high prevalence of osteoporosis and high cancer incidence rates. Imaging modalities such as CT and MRI play an important role in determining the benignity or malignancy of vertebral fractures. The widespread availability, speed and affordability of CT have led to its frequent use in the evaluation of vertebral fractures. In several recent studies, CT radiomics has shown promising results in the evaluation of vertebral fractures6,9, including the ability to successfully differentiate malignant from acute benign compression fractures6. These findings suggest that CT radiomics may provide an alternative diagnostic approach to determine the etiology of vertebral fractures. In that study, however, the fractured vertebrae were segmented manually, limiting the routine clinical implementation of the proposed prediction model. We hypothesized that this limitation could be overcome by automated segmentation, potentially leading to more successful implementation of radiomics in clinical practice.

Therefore, in this study, we aimed to develop and validate an automated algorithm for segmentation of fractured vertebral bodies on CT. Additionally, to evaluate the applicability of automated algorithm for use in radiomics, the algorithm was compared with the human expert segmentation for the prediction performance of a radiomics model to differentiate between acute benign and malignant compression fractures.

This retrospective study was approved by the Institutional Review Board of the Asan Medical Center (approval no. 2019-0134), Institutional Review Board of the Seoul National University Bundang Hospital (no. B-2008/628-109) and Institutional Review Board of the Inha University Hospital (no. 2020-08-018), and the requirement to obtain informed patient consent was waived. All methods were performed in accordance with the relevant guidelines and regulations.

This study included patients who (a) underwent spine CT for acute benign or malignant vertebral compression fractures in the thoracic and lumbar vertebrae between January 2015 and April 2020, and (b) underwent MRI within 6 weeks of CT examination. Acute benign compression fractures were defined as traumatic or osteoporotic fractures with abrupt onset of back pain of less than 6 weeks27, whereas malignant fractures were defined as fractures replaced or infiltrated by tumor tissue28. In addition, chronic fractures were defined as old, healed benign compression fractures without bone marrow edema on MRI. Patient exclusion criteria are shown in the flow diagram (Fig. 1).

The benignity or malignancy and the acuity or chronicity of fracture was determined by a musculoskeletal radiologist with 10 years of experience in spine imaging, based on MRI performed within 6 weeks of CT examination, and, if available, follow-up imaging or pathologic confirmation of tissue samples obtained surgically or on percutaneous biopsies.

An overview of the development of the CNN and its detailed architecture are presented in Fig. 2. The proposed fractured vertebral body segmentation method was composed of two steps: vertebral detection and segmentation.

YoloV3 is a one-stage detector that uses multi-scaled feature maps and predefined anchor boxes to rapidly and accurately predict localization and class of bounding boxes35. The baseline consisted of the YoloV3 framework35, followed by the modifications that included (a) application of dense connection and separable convolution to the yolo block, and (b) effective reduction of the scale layer through data augmentation and optimization of the anchor box and grid size for vertebrae. These enabled efficient improvements of accuracy and rapid ROI extraction. Each vertebral ROI extracted from the coronal MIP image was used as a limit to generate a sagittal MIP image. Vertebral ROIs were extracted from the sagittal MIP images in the same manner as in the coronal plane. Vertebral VOIs were generated using the minima and maxima for the x, y, and z coordinates of each ROI extracted from the two planes of MIP images.

Because severe bone destruction in some cases made it difficult to determine the total morphology of each vertebra, segmentation was first performed in the sagittal plane, followed by the axial plane. Thin-slab MIPs were generated from continuous slices, followed by propagation of reduced segmentation areas to adjacent areas to improve segmentation performance. Thin-slab MIPs compensated for the partial loss or broken regions by merging information from the n-th adjacent slice. Propagation maintained the topologic characteristics of the overall fractured vertebral body based on the linear characteristics of the adjacent regions on CT images. The results of segmentation in the sagittal plane were used to reconstruct images in the axial plane. At this time, the CNN segmentation prediction area was reduced using the distance map in order to solve the over-segmentation problem caused by thin-slab MIP generation.

Our network was based on U-Net framework36. The CNN consisted of encoding and decoding paths, and the performance was improved through EfficientNet37 in the encoding path and Attention U-Net38 in the decoding path. Application of the compound scaling method to the encoding path of the proposed CNN architecture reduced calculation costs and improved accuracy. In the decoding path, the attention block emphasized important features of the vertebrae, progressively suppressing the feature response to the background area. The numbers of parameters were effectively reduced in the encoding and decoding paths, improving the performance. A total of five resolution steps were used in all experiments.

Binary Cross Entropy (BCE) and Dice Loss39 were performed to minimize background bias in vertebral segmentation results. In addition, propagation loss was used to compensate for partially lost or broken regions. The overall loss function can be defined as:

Then, fivefold cross-validation was performed using least absolute shrinkage and selection operator (LASSO) regression with penalty parameter tuning to select significant radiomics features with non-zero coefficients that can predict malignancy of vertebral fracture. Finally, a radiomics model was constructed from linear combinations of features weighted by LASSO coefficients.

The diagnostic performance of the radiomics prediction model was compared between the automated and human expert segmentations on the internal and external test sets. The performance of the model was evaluated using the area under the receiver operating characteristics curve (AUC) and compared using the Delong method.

Subgroup analysis showed that the algorithm achieved the highest performance for chronic benign fractures, followed by acute benign fractures and malignant fractures, with statistically significant differences, except for the CSA error of the internal test set.

In this study, we developed and validated an automated algorithm for segmentation of fractured vertebral bodies on CT. The algorithm achieved high agreement with the human expert segmentation on two independent test sets. In addition, the automated and the human expert segmentation methods were compared for the prediction performance of a radiomics model to differentiate acute benign and malignant compression fractures, and the two segmentation methods showed comparable discrimination performance and accuracy, indicating the applicability of the proposed algorithm for use in radiomics.

Automated vertebral segmentation is needed for many purposes, including diagnosis and treatment planning. However, as Rizzo et al. mentioned, there is no universal segmentation algorithm for all applications and purposes43. We sought to develop an algorithm for subsequent use in radiomics analysis to differentiate acute benign and malignant compression fractures. While several previous works on automated segmentation only evaluated the reproducibility of radiomics features (i.e., correlations between automated and manual segmentation) as one measure of segmentation performance5,44, the extent of feature reproducibility may not directly translate into the performance of a radiomics model. Therefore, automated segmentation was compared with human expert segmentation in the performance of a radiomics model, which was constructed with features robust against segmentation variability, and the algorithm and the human experts showed comparable performances. These results suggest the applicability of the automated algorithm for use in a radiomics prediction model.

In conclusion, we developed and validated an automated algorithm for segmentation of fractured vertebral bodies on CT. The automated algorithm showed comparable performance to the human expert segmentation in a CT radiomics model to predict fracture malignancy, which may enable more practical clinical utilization of radiomics.

3a8082e126