Correlation between whole skeleton dual energy CT calcium-subtracted attenuation and bone marrow infiltration in multiple myeloma

Highlights • Quantification of whole skeleton calcium-subtracted attenuation with dual energy CT is feasible.• Whole skeleton calcium-subtracted attenuation correlates with the degree of marrow infiltration by plasma cells on bone marrow biopsy.• Whole skeleton calcium-subtracted attenuation provides complementary information to the detection of osteolytic bone lesions.


Introduction
Multiple myeloma is a bone marrow haemopathy characterised by clonal plasma cells, causing pathological fractures, anaemia, recurrent infections, hypercalcaemia, and renal failure [1]. 140,000 new cases are diagnosed worldwide each year [2]. The 2019 International Myeloma Working Group (IMWG) imaging guidelines currently recommend whole body CT as a first line diagnostic test for suspected myeloma [3]. The presence of one or more osteolytic lesion on CT is a myeloma defining event [4]. However, false positive lesions may occur due to focal areas of yellow marrow, vertebral end plate degenerative changes or haemangiomas [5]. Marrow infiltration by plasma cells may be present prior to the development of visible osteolytic lesions on CT. Thus, assessment of skeletal marrow infiltration may detect disease or its Abbreviations: IMWG, International Myeloma Working Group. progression at an earlier time point than current practice.
While CT attenuation assessment is not currently recommended for diagnosis by the IMWG guidelines, dual energy CT with reconstruction of calcium-subtracted attenuation maps offers an opportunity to quantify the degree of bone marrow infiltration. For example, a previous study of 53 patients with MRI of the axial skeleton as a reference standard showed differences in calcium-subtracted attenuation between normal, focal and diffuse imaging patterns with mean calcium attenuation (Hounsfield unit, HU) of − 66HU, +3HU and − 13HU, respectively [6].
To date, dual energy CT studies in myeloma patients have focussed on placing focal regions-of-interest within MRI correlated-CT bone lesions and/or within selected vertebrae or selected areas within the pelvis [6][7][8][9]. Focal lesional attenuation at a high level of calcium suppression has also been shown to predict for 18F-FDG PET metabolic activity, with higher attenuation noted in metabolically active vs. non-active lesions [10]. More recently, the feasibility of artificial intelligence tools to segment the thoracolumbar spine and quantify attenuation have been explored [11]. Here, quantified CT of the non-fatty portion of spinal bone marrow predicted pelvic bone marrow infiltration after adjustment for bone mineral density with r = 0.46 [11]. However, tools for objective assessment of the whole skeleton are still lacking.
We hypothesized that dual energy CT derived calcium-subtracted attenuation of the whole imaged skeleton will reflect the degree of bone marrow infiltration, providing complementary information to the presence of focal osteolytic lesions or regional marrow assessment.
Thus, we aimed to assess the feasibility of whole skeleton quantification of calcium-subtracted attenuation; and to assess its correlation with bone marrow plasma cell infiltration percentage from biopsy. Secondarily, we aimed to correlate whole skeleton calcium-subtracted HU attenuation values with other clinical variables (haemoglobin level and age); and to compare whole skeleton calcium-subtracted HU attenuation values to those derived from focal marrow evaluation and osteolytic lesions.

Material and methods
Institutional approval was obtained for this study and patients provided informed consent for this analysis.

Patients
Consecutive patients with suspected or proven newly-diagnosed or relapsed myeloma, who were unable to undergo whole body MRI as per our usual clinical practice (for reasons including claustrophobia (n = 5); whole body MRI not tolerated due to symptoms (n = 5); non-MRI compatible implant (n = 1); physician triage/choice (n = 10)) prior to treatment, and who attended for a dual energy CT between 01 February 2018 and 01 April 2021 were included. There were no predefined exclusions. The study flowchart is summarised in Fig. 1.

Dual energy CT acquisition, image reconstruction and post-processing
All patients underwent a supine non-contrast dual energy CT from the skull vertex to proximal tibia on a third generation dual source CT scanner (Somatom Force, Siemens Healthineers, Forchheim, Germany) with the following acquisition parameters: tube voltages 90 kV (A) / 150 kV with tin filter (B); collimation, 128x 0.6 mm; pitch, 0.6; rotation time, 0.5 s; automated attenuation based tube current modulation (Care dose 4D; Siemens Healthineers, Forchheim, Germany); and quality reference tube current-time product, 1.6:1.
Axial images (slice thickness 1.0 mm; 0.75 mm increment; 500 mm field-of-view; soft tissue kernel QR40; advanced model iterative reconstruction (ADMIRE) level 3) were reconstructed for 90kVp and 150kVp datasets. Axial weighted-average images with equivalent contrast to single energy 120kVp images were also derived from these datasets with a 50:50 mixing ratio for viewing purposes as per our clinical practice.
For viewing purposes, dual energy CT images were also displayed as weighted-average CT images with a colour-coded virtual calciumsubtracted CT overlay using the bone marrow settings. Representative patient examples are shown in Fig. 2. The mean ± SD CT dose index was 7.81 ± 1.96 mGy.cm and the mean ± SD dose-length product was 1017 ± 322 mGy.cm for this cohort.

Segmentation and generation of whole skeleton calcium-subtracted attenuation values
Anonymised images were exported for further post-processing. Whole skeleton segmentation and quantification was undertaken using an in-house developed semi-automatic pipeline written in Python within the Pytorch environment. The pipeline is illustrated in Fig. 3. The weighted-average 120 kVp equivalent CT images were reconstructed with bone windows (W: 1800 L:400) and normalised with the median HU value. Following stepwise assessment for the optimal thresholding value above 100HU for contouring the skeleton, a fixed thresholding value of 120 HU was then applied to generate an initial contour for further post-processing. Next, to fill in this contour and to exclude nonskeletal structures (i.e. brain, spinal cord, spinal canal), a Chan-Vese morphological operation (morphological active contours without edges) was applied [12].
The Chan-Vese morphological operation iteratively minimises the energy function F(c 1 ,c 2 ,C) by the following equation: where, µ = 1 is the penalty on the total length of the boundary of the segmented region (C); while λ 1 ,λ 2 are two weighting parameters affecting the uniformity of the inner contour (inside(C)) and outer contour region (outside(C)). Here, both were set to 1 for binary mask generation (i.e., c 1 = c 2 = 1).
With this operation, the brain, spinal cord, spinal canal, and cerebrospinal fluid were excluded as the mean pixel value of the inner contour region is substantially different from the outer contour region. Two iterations of the Chan-Vese morphological operation were used initially for the whole skeleton. The mask was then reviewed by a radiologist, and, where necessary, further slice numbers were inputted to enable further filling in ('erosion') of the skeleton of selected images (e.g., of the femur or humerus) to ensure that the skeletal bone marrow mask was complete and representative. Any non-skeletal structures included in the process, e.g. hip prosthesis, were removed by further HU thresholding ± cropping. The resultant skeletal bone marrow mask was then applied directly to the corresponding calcium-subtracted attenuation parametric map to extract a histogram of bone marrow HU values for the whole skeleton.

Generation of regional marrow and lesion calcium-subtracted attenuation values
In all patients, using the vendor dedicated dual energy CT software on Syngo.via (VB30; Siemens Healthineers, Forchheim, Germany) a standard circular region-of-interest was placed within the L3 vertebra and right iliac bone (excluding any focal osteolytic lesion, if present) to generate calcium-subtracted attenuation values. Mean and standard deviation calcium-subtracted Hounsfield unit values were generated by the software and recorded.
In patients with focal osteolytic bone lesions, up to 5 focal skeletal lesions (>5mm) were identified on the weighted-average images and corresponding calcium-subtracted images. A circular region-of-interest was placed within the boundaries of each lesion. Again, mean and standard deviation calcium-subtracted Hounsfield unit values for each lesion generated by the software were recorded.

Clinical management and follow up
This was undertaken as per usual institutional practice. Patients were staged using the International Staging System. Bone marrow biopsy was performed where this directed further clinical management; sampling was from the iliac crest. Blood tests included a full blood count, urea and electrolytes, and bone biochemistry.

Statistical analysis
Descriptive statistics were performed for patient and disease characteristics. Spearman's rank correlation assessed associations between whole skeleton calcium-subtracted attenuation and biopsy derived marrow plasma cell infiltration level, blood haemoglobin level, and age, respectively. Wilcoxon signed-rank test was performed to compare grouped lesion calcium-subtracted attenuation, grouped regional and whole skeleton calcium-subtracted attenuation.  Table 1.

Whole skeleton segmentation and calcium-subtracted attenuation
Segmentation was feasible in all patients. In 17/21 (81%) patients, no further adjustments were required to the skeletal marrow mask generated by the segmentation pipeline. In 3/21 (14%) patients, metal prostheses and a loop of hyperdense bowel required additional postprocessing for removal. Image processing time ranged from 5 to 8 min, depending on the need for additional adjustments. Histogram values derived from whole skeleton segmentation for each patient are summarised in Table 2.

Comparison of whole skeleton with regional marrow calciumsubtracted attenuation values
Compared with regional average calcium-subtracted attenuation derived from L3 and the iliac crest, whole skeleton average calciumsubtracted attenuation had a lower median value (-59.9 HU vs. − 51.3 HU; p < 0.001) highlighting regional differences in marrow values.

Comparison of whole skeleton with lesion calcium-subtracted attenuation values
Compared with lesion average calcium-subtracted attenuation, whole skeleton average calcium-subtracted attenuation had a lower median value (-56.8 HU vs. − 11.8 HU; p < 0.001) indicating the complementary nature of these measurements.

Discussion
There is still an outstanding clinical need for earlier and more objective assessment of skeletal infiltration by plasma cells in multiple myeloma. In addition to osteolytic lesions arising from bone destruction, osteoporosis has been reported in patients with multiple myeloma undergoing CT [13] indicating a more diffuse process underlying CT findings. To date, assessment in clinical practice has been limited by the lack of whole skeleton assessment tools, with published studies reporting on regional analysis from focal region-of-interest analysis of the lumbar spine or pelvis.
In this cohort of patients, we found that whole skeleton segmentation was feasible using an in-house segmentation and quantification pipeline, and that whole skeleton calcium-subtracted attenuation values correlated positively with biopsy plasma cell infiltration level, yet, were distinct from regional marrow and lesional calcium-subtracted attenuation values, indicating their complementary nature for diagnosis and ongoing disease assessment. The processing time of 5-8 min was acceptable for clinical practice. Further adjustment of the segmentation mask was required in 10% of patients in our cohort, predominantly to remove metal prostheses. An advantage of this pipeline is that it would also be applicable to other dual energy CT methods including those integrated into PET/CT scanners and with calcium-subtracted mapping capabilities.
Previous dual energy CT studies have indicated that visual analysis may detect bone marrow infiltration with good diagnostic performance, but this will always be subject to observer variation. Calcium-subtracted cut-off values of − 44.9HU have been quoted for the presence of bone marrow infiltration (using MRI as reference standard) with a sensitivity and specificity of 93% [7]. However, being derived from focal region-ofinterest analysis rather than the whole skeleton, such analysis may not capture the heterogenous distribution of plasma cells associated with myeloma throughout the skeleton. Similarly, cut-off values for lytic lesions of − 3 HU on CT have been quoted [8]. Our quantitative values were in a comparable range to these previously quoted values but clearly, whole skeleton analysis provides more objective evidence of the global extent of marrow infiltration.
To date, studies have highlighted a role for MRI rather than CT for the evaluation of marrow infiltration in multiple myeloma due to MRI's higher sensitivity for infiltrative disease compared to standard CT [14]. In addition to the assessment of different patterns of bone marrow involvement (normal, focal, homogeneous diffuse infiltration, combined focal and diffuse infiltration and variegated or 'salt and pepper') [15], quantitative assessment of skeletal T1-weighted gradient echo Dixon

Creatinine level
Normal range (48-84 μmol/L)  Abbreviation: MGUS = monoclonal gammopathy of unknown significance; SD = standard deviation derived fat-signal fraction and diffusion-weighted MRI derived apparent diffusion co-efficient have advocated e.g., of focal lesions for therapy assessment in clinical trial settings [16][17][18][19]. In terms of marrow assessment, fat-signal fraction within the spine (L1-L3) has been shown to be lower in symptomatic versus asymptomatic myeloma [20,21]; and lower in symptomatic versus healthy controls [22], assumed to indicate replacement of marrow fat cells by plasma cell infiltration. Conversely, apparent diffusion co-efficient derived from ROI analysis of the thoracic and lumbar spine have been shown to be higher in myeloma patients with diffuse infiltration compared to asymptomatic myeloma patients and healthy controls [22]. As with quantitative dual energy CT and 18F-FDG PET, a challenge remains of differentiation of infiltrative disease from hyperplastic haematopoietic bone marrow, with lower therapeutic assessment performance in patients with anaemia [23]. Combining assessment of fat-signal fraction and apparent diffusion coefficient with anatomical appearances may potentially mitigate this [24]. For patients who cannot undergo whole body MRI, dual energy CT provides an alternative. Nevertheless, there are limitations to this study. First, the sample size for this feasibility study was small. However, the findings remain encouraging. Second, although there was a positive correlation with plasma cell infiltration, a number of additional factors are known to contribute to skeletal calcium-subtracted attenuation values including the level of other marrow components, e.g., haematopoietic cells; this was suggested indirectly by the negative correlation with haemoglobin level in our study. Third, although we employed a phantom-optimized dual energy CT protocol balancing radiation dose to sufficient image quality for accurate quantification, this cannot be deemed a low-dose CT protocol. Thus, the benefits of quantification have to be balanced against a radiation burden to the patient. Further exploration of lower dose techniques including low dose dual energy CT with higher levels of modelled iterative reconstruction or dual layer detector dual energy CT is required. Fourth, our cohort were preselected, and underwent dual energy CT rather than whole body MRI thus no comparisons could be undertaken between both modalities in this cohort. Finally, day-to-day test-retest analysis was not performed in this cohort as only one CT examination was performed but would be valuable information to obtain in the future.
In conclusion, whole skeleton segmentation was feasible. Calciumsubtracted attenuation was associated positively with the degree of marrow infiltration, differed from lesional values, and may provide a more objective measure of overall disease burden than current practice.