A CT Scan Combined with Deep Learning Can Also Detect Osteoporosis

Introduction: Osteoporosis is a bone metabolic disorder characterized by a decrease in bone mineral density (BMD) and strength, which increases the risk of low back pain, disc degeneration, or vertebral fracture.
Osteoporosis is a common and frequent disease in the aging population.
Osteoporosis is a bone metabolic disease characterized by a decrease in bone mineral density (BMD) and strength, which increases the risk of low back pain, disc degeneration, or vertebral fracture.
Therefore, early diagnosis of osteoporosis is very important to prevent disease progression.
Currently, conventional methods for evaluating osteoporosis include dual-energy X-ray absorption (DXA), quantitative computed tomography (QCT), and quantitative ultrasound (QUS), as well as emerging imaging technologies such as dual-energy spectrum CT, 1H-MRS, and positron emission tomography (PET).
BMD measurement is a reliable and ideal method for early diagnosis of osteoporosis.
DXA is a commonly used tool for measuring bone density in the spine.
However, DXA was unable to eliminate the effect of cortical, hyperosteogenic, and sclerosis on BMD measurements, which may underestimate the actual loss of bone mass.
QCT is a recognized 3D method for bone mineral density assessment.
Several studies have shown that the detection rate of QCT for osteoporosis is significantly higher than that of DXA.
However, compared with DXA, QCT is difficult to be widely used in clinic due to its high radiation dose and complex post-processing.
Each year, patients with other indications, such as urinary or digestive disorders, undergo a CT scan of part or all of the spine.
These CT scans can be used for opportunistic screening for osteoporosis without additional radiation exposure and expense.
At present, deep learning has been increasingly used in medical image analysis.
Recently, published in the European journal of Radiology of a study to explore the deep learning application in patients with primary osteoporosis, and developed a convolutional neural network based on depth (deep convolutional neural network, the DCNN) automatic CT image segmentation and bone mineral density and vertebral bodies (ipads mineral density, BMD) calculation method, designed to automatically from accuracy assessment methods lumbar vertebral BMD and calculate the positioning performance.
In this study, 1449 patients who underwent CT scans of the spine or abdomen for other indications between March 2018 and May 2020 were retrospectively selected for validation and analysis.
All data were obtained from three different CT providers.
Of these, 586 were used for training and 863 for validation.
The full convolutional neural network U-NET was used to segment the vertebral body automatically.
The manually drawn vertebral body area was used as the base map for comparison.
Bone mineral density was calculated using Densenet-121 convolutional neural network.
Quantitative computed tomography (QCT) post-processing values were used as the criteria for the analysis.
According to different CT providers, all test cases were divided into the following three test queues: test set 1 (n = 463), test set 2 (n = 200), and test set 3 (n = 200).
There was a good correlation between automatic segmentation and the four manually segmented lumbar spine (L1-L4) : the minimum average DICE coefficients of the three test sets were 0.823, 0.786, and 0.782, respectively.
For the test set from different suppliers, the mean BMD calculated by automatic regression showed a high correlation with the results obtained from QCT (R>
0.98) and consistency.
Fig. Visual comparison of automatic and manual segmentation results.
From top to bottom are the segmentation results of test set 1, 2 and 3 respectively.
CT sagittal images, manual segmentation and automatic segmentation are shown from left to right.
Red, green, yellow, and blue represent the vertebral bodies of L1, L2, L3, and L4, respectively
The correlation between manual and automatic segmentation was very good, with DSCs exceeding 0.90 for all four lumbar vertebrae.
This study shows that the DCNN-based method can accurately segment the lumbar vertebrae and automatically calculate the bone mineral density, making it an effective tool for clinicians to screen for opportunistic osteoporosis.