A quantitative assessment of dual energy computed tomography‐based material decomposition for imaging bone marrow edema associated with acute knee injury

Abstract

This study developed methods to quantify and improve the accuracy of dual-energy CT (DECT)-based bone marrow edema imaging using a clinical CT system. Objectives were: (a) to quantitatively compare DECT with gold-standard, fluid-sensitive MRI for imaging of edema-like marrow signal intensity (EMSI) and (b) to identify image analysis parameters that improve delineation of EMSI associated with acute knee injury on DECT images. Methods: DECT images from ten participants with acute knee injury were decomposed into estimated fractions of bone, healthy marrow, and edema based on energy-dependent differences in tissue attenuation. Fluid-sensitive MR images were registered to DECT for quantitative, voxel-by-voxel comparison between the two modalities. An optimization scheme was developed to find attenuation coefficients for healthy marrow and edema that improved EMSI delineation, compared to MRI. DECT method accuracy was evaluated by measuring dice coefficients, mutual information, and normalized cross correlation between the DECT result and registered MRI. Results: When applying the optimized three-material decomposition method, dice coefficients for EMSI identified through DECT vs MRI were 0.32 at the tibia and 0.13 at the femur. Optimization of attenuation coefficients improved dice coefficient, mutual information, and cross-correlation between DECT and gold-standard MRI by 48%-107% compared to three-material decomposition using nonoptimized parameters, and improved mutual information and cross-correlation by 39%-58% compared to the manufacturer-provided two-material decomposition. Conclusions: This study quantitatively evaluated the performance of DECT in imaging knee injuryassociated EMSI and identified a method to optimize DECT-based visualization of complex tissues (marrow and edema) whose attenuation parameters cannot be easily characterized. Further studies are needed to improve DECT-based EMSI imaging at the femur.

Quantification of spinal bone marrow fat fraction using three‐material decomposition technique on dual‐energy CT: A phantom study

“…This study demonstrates that this limitation can be overcome by an imagedomain approach. Our method further differentiates from previously proposed image-domain three-material decomposition methods 10,[16][17][18] in that the aforementioned methods rely on custom-made calibration phantoms and patient-specific calibration process, while our approach takes advantage of the mass fraction of hydroxyapatite readily available from the Revolution CT scanner. While this limits our approach to certain scanner models, it avoids variabilities and uncertainties introduced by the custom-designed calibration phantoms and the patient-specific calibration process.…”

“…This technique decomposes bone minerals, water, and fat from a pair of high and low monochromatic energy images, based on their attenuation coefficients at the specific energies. In contrary to other image-domain approaches, 10,[16][17][18] our approach takes advantage of the mass fraction of hydroxyapatite readily available from the commercial dual-energy CT scanner used in this study, hence no longer requires any patient-specific calibration. The technique is described in detail in the next section, followed by a phantom study, to evaluate its efficacy.…”

MsgeCNN: Multiscale geometric embedded convolutional neural network for ONFH segmentation and grading

“…With the development of artificial intelligence and computer vision technology, deep learning methods have shown huge application potential in the computer-aided diagnosis (CAD) of orthopedic diseases. These methods have been applied for a variety of orthopedic diseases, including knee joint, 9 bone cancer, 10 and so forth. For example, Nguyen et al 9 proposed a semisupervised learning (SSL) method for knee osteoarthritis (OA) image classification using unlabeled data.…”

“…These methods have been applied for a variety of orthopedic diseases, including knee joint, 9 bone cancer, 10 and so forth. For example, Nguyen et al 9 proposed a semisupervised learning (SSL) method for knee osteoarthritis (OA) image classification using unlabeled data. Since image classification networks can only give image-level diagnostic results, Tiulpin et al 11 extracted attention maps of radiological features, which can show image regions that mainly affect network decisions.…”

“…These methods have been applied for a variety of orthopedic diseases, including knee joint, 9 bone cancer, 10 and so forth. For example, Nguyen et al 9 . proposed a semi‐supervised learning (SSL) method for knee osteoarthritis (OA) image classification using unlabeled data.…”