An in-depth review and analysis of deep learning methods and applications in spinal cord imaging

Should Spinal MRI Be Routinely Performed in Patients With Clinically Isolated Optic Neuritis?

Quantitative spinal cord imaging: Early ALS diagnosis and monitoring of disease progression

Objective The automation of brachytherapy is the direction of future development. This article retrospectively studied the application of deep learning in brachytherapy of cervical cancer and clarified the status quo of development. Method This survey reviewed the application of machine learning and deep learning in brachytherapy for cervical cancer in the past 10 years. The survey retrieved and reviewed electronic journal articles in scientific databases such as Google Scholar and IEEE. The three sets of keywords used 1) deep learning, brachytherapy, 2) machine learning, brachytherapy, 3) automation, brachytherapy. Results Through research on the application of deep learning in brachytherapy, it is found that the U-net model is basically based on convolutional neural networks or some attention mechanisms are added to it, and it is applied to brachytherapy of prostate or cervical cancer. The automatic segmentation and reconstruction of the mid-source applicator (interpolation needle), target area delineation, optimization in the treatment planning system and dose calculation have achieved good results, proving that deep learning can be applied to the clinical treatment of brachytherapy. Conclusion The research on the application of deep learning in brachytherapy confirmed that deep learning can effectively promote the development of brachytherapy.

The spinal cord plays a critical role in the nervous system. Structurally, morphological changes in the spinal cord directly reflect nerve fiber tract reorganization and may indicate underlying pathological alterations in gray and white matter. Structural remodeling of the spinal cord underlies its functional impairments. This presents a range of clinical manifestations, such as partial or complete loss of voluntary motor control, altered reflexes, abnormal muscle tone, muscular atrophy, and sensory deficits. Spinal cord imaging provides researchers with a direct approach to explore morphological alterations. Innovations in artificial intelligence and automated processing software now enable quantitative analysis of volumetric structural changes. Currently, spinal cord imaging markers play an increasingly prominent role in tracking disease progression and elucidating pathogenic mechanisms across neurological disorders. Despite these advances, these markers have yet to fully capture the complexity of the clinical disease presentation. This article aims to systematically review recent advances in spinal cord imaging and pathology across neurodegenerative and neuroinflammatory diseases. Furthermore, it proposes future research directions for spinal cord studies leveraging current imaging analysis techniques. This review comprehensively summarizes existing research on spinal cord imaging and pathological features in varied neurological diseases. It focuses on the relationships between spinal cord imaging markers and clinical features, disease severity, predictive value in clinical contexts and discusses future research directions for spinal cord quantification in neurological disorders.

MR imaging plays an important role in diagnosing MS and other related inflammatory diseases; however, imaging of the spinal cord is still challenging. We hypothesized that a 3D double inversion recovery sequence for cervical spinal cord imaging would be more sensitive in detecting inflammatory lesions than a conventional 2D T2-weighted TSE sequence at 3T. On a 3T MR imaging scanner, we examined 30 patients with suspected or established MS (MS, n = 16; clinically isolated syndrome, n = 12; isolated myelitis, n = 2) and 10 healthy controls. Newly developed 3D double inversion recovery and conventional 2D axial and sagittal T2-weighted TSE images of the cervical spinal cord were acquired. Two blinded neuroradiologists independently assessed the scans in pseudorandomized order for lesion numbers and rated lesion visibility and overall image quality on 5-point scales. A subsequent consensus reading delivered definite lesion counts. Standardized contrast-to-noise ratios were calculated in representative lesions of each patient. Overall, 28% more lesions could be detected with 3D double inversion recovery than with conventional T2WI (119 versus 93, P < .002). On average, the standardized contrast-to-noise ratio was significantly higher (P < .001) in double inversion recovery than in T2WI. Lesion visibility was rated significantly higher (P < .001) in double inversion recovery compared with T2WI despite lower image quality. The novel 3D double inversion recovery sequence allowed better detection of lesions in MS and related inflammatory diseases of the cervical spinal cord, compared with conventional 2D T2WI.

To establish a novel approach for fast high-resolution spinal cord (SC) imaging using averaged magnetization inversion recovery acquisitions (AMIRA). The AMIRA concept is based on an inversion recovery (IR) prepared, segmented, and time-limited cine balanced steady state free precession sequence. Typically, for the fastest SC imaging without any signal averaging, eight consecutive images in time with an in-plane resolution of 0.67 × 0.67 mm2 and 6 mm to 8 mm slice thickness are acquired in 51 s. AMIRA does not require parallel acquisition techniques. AMIRA measures eight images of remarkable tissue contrast variation between spinal cord gray (GM) and white matter (WM) and cerebrospinal fluid (CSF). Following the AMIRA concept, averaging the first IR contrast images not only improves the signal-to-noise ratio but also offers a surprising enhancement of the contrast-to-noise ratio between GM and WM, whereas averaging the last images considerably improves the contrast-to-noise ratio between WM and CSF. These observations are supported by quantitative data. The AMIRA concept provides 2D spinal cord imaging with multiple tissue contrasts and enhanced contrast-to-noise ratios with a typical 0.67 × 0.67 mm2 in-plane resolution and a slice thickness between 4 mm and 8 mm acquired in only 1 to 2 min per slice. Magn Reson Med 79:1870-1881, 2018. © 2017 International Society for Magnetic Resonance in Medicine.

The rapid development of Internet technology has made the whole society enter the era of big data. In recent years, the development trend of artificial intelligence and machine learning has also risen sharply. Informatization has become an important feature of the current era. As an indispensable common information carrier, images not only facilitate people’s communication, but also promote the development of deep learning processing image technology. Based on this, this paper analyzes the application of computer deep learning in image processing. Firstly, the deep learning is summarized, its concept and origin are briefly introduced, and then the technical classification, development process and processing purpose of image processing are expounded. Finally, the application of computer deep learning in four aspects is analyzed in detail in image recognition, image denoising, image classification and image enhancement, and it has certain significance for promoting the research and application of deep learning.

Amyotrophic lateral sclerosis (ALS) is the most common adult onset motor neuron disease with no effective disease modifying therapies at present. Spinal cord degeneration is a hallmark feature of ALS, highlighted in the earliest descriptions of the disease by Lockhart Clarke and Jean-Martin Charcot. The anterior horns and corticospinal tracts are invariably affected in ALS, but up to recently it has been notoriously challenging to detect and characterize spinal pathology in vivo. With recent technological advances, spinal imaging now offers unique opportunities to appraise lower motor neuron degeneration, sensory involvement, metabolic alterations, and interneuron pathology in ALS. Quantitative spinal imaging in ALS has now been used in cross-sectional and longitudinal study designs, applied to presymptomatic mutation carriers, and utilized in machine learning applications. Despite its enormous clinical and academic potential, a number of physiological, technological, and methodological challenges limit the routine use of computational spinal imaging in ALS. In this review, we provide a comprehensive overview of emerging spinal cord imaging methods and discuss their advantages, drawbacks, and biomarker potential in clinical applications, clinical trial settings, monitoring, and prognostic roles.

Despite considerable involvement of the spinal cord in amyotrophic lateral sclerosis (ALS), current biomarker research is primarily centred on brain imaging and CSF proteomics. In clinical practice, spinal cord imaging in ALS is performed primarily to rule out alternative conditions in the diagnostic phase of the disease. Quantitative spinal cord imaging has traditionally been regarded as challenging, as it requires high spatial resolution while minimizing partial volume effects, physiological motion and susceptibility distortions. In recent years however, as acquisition and post-processing methods have been perfected, a number of exciting and promising quantitative spinal imaging and electrophysiology techniques have been developed. We performed a systematic review of the trends, methodologies, limitations and conclusions of recent spinal cord studies in ALS to explore the diagnostic and prognostic potential of spinal markers. Novel corrective techniques for quantitative spinal cord imaging are systematically reviewed. Recent findings demonstrate that imaging techniques previously used in brain imaging, such as diffusion tensor, functional and metabolic imaging can now be successfully applied to the human spinal cord. Optimized electrophysiological approaches make the non-invasive assessment of corticospinal pathways possible, and multimodal spinal techniques are likely to increase the specificity and sensitivity of proposed spinal markers. In conclusion, spinal cord imaging is an emerging area of ALS biomarker research. Novel quantitative spinal modalities have already been successfully used in ALS animal models and have the potential for development into sensitive ALS biomarkers in humans.

Long-term and non-narcotic hemodynamic imaging is indispensable for observing factual physiological information of the spinal cord. Unfortunately, achieving label-free, high-resolution, and widefield spinal cord imaging for mice under freely moving conditions is challenging. In this study, we developed a miniaturized photoacoustic microscope along with a corresponding photoacoustic spinal window to realize high-resolution, multi-segmental hemodynamic imaging of the spinal cord for freely moving mice. The microscope has an outer size of 32 mm × 23 mm × 10 mm, a weight of 5.8 g, and a 4.4 µm lateral resolution within an effective field of view (FOV) of 2.6 mm × 1.8 mm. To eliminate the off-focus phenomena during spinal imaging, the microscope is equipped with a miniature motor to adapt the focal plane. Besides, the microscope is slidable along a customized rail on the window to expand the FOV. We evaluated the stability of the microscope and analyzed vascular images of the spinal cord under various physiological states. The results suggest that the microscope is capable of performing stable, multi-segmental spinal cord imaging in freely moving mice, offering new insights into spinal cord hemodynamics and neurovascular coupling research.

A first-ever spinal cord imaging meeting was sponsored by the International Spinal Research Trust and the Wings for Life Foundation with the aim of identifying the current state-of-the-art of spinal cord imaging, the current greatest challenges, and greatest needs for future development. This meeting was attended by a small group of invited experts spanning all aspects of spinal cord imaging from basic research to clinical practice. The greatest current challenges for spinal cord imaging were identified as arising from the imaging environment itself; difficult imaging environment created by the bone surrounding the spinal canal, physiological motion of the cord and adjacent tissues, and small crosssectional dimensions of the spinal cord, exacerbated by metallic implants often present in injured patients. Challenges were also identified as a result of a lack of “critical mass” of researchers taking on the development of spinal cord imaging, affecting both the rate of progress in the field, and the demand for equipment and software to manufacturers to produce the necessary tools. Here we define the current state-of-the-art of spinal cord imaging, discuss the underlying theory and challenges, and present the evidence for the current and potential power of these methods. In two review papers (part I and part II), we propose that the challenges can be overcome with advances in methods, improving availability and effectiveness of methods, and linking existing researchers to create the necessary scientific and clinical network to advance the rate of progress and impact of the research.

The current state-of-the-art of spinal cord imaging: Methods

Intensity inhomogeneity correction is one of the challenges in Magnetic Resonance Imaging (MRI). Correction of intensity inhomogeneity for spinal cord images is more difficult than brain images because of the small size of homogeneous regions such as gray and white matter structures in a spinal cord. In this paper we present a pixel based correction method called Improved Adaptive Gamma Correction Method (IAGCM). Several qualitative and quantitative evaluations were tested on spinal cord images on seven healthy and three subjects with spinal cord injury. Contrast to Noise Ratio (CNR) and Signal to Noise Ratio (SNR) for different parameters of “p” were considered as quantitative evaluation. Qualitative evaluations were performed by scoring of images by an independent board certified neuroradiologist on images applied with pseudo color transformation. Overall the results showed improvement in the CNR of the corrected images. In conclusion, the proposed IAGCM method is a simple method to implement and performed well in improving the image quality of spinal cord images.

We illustrate the contribution of high-frequency linear abdominal transducers in the prenatal US examination of the spinal cord. After birth, such transducers are commonly used in US examination of the spinal cord. During the third trimester of gestation, the fetal spine is commonly facing anteriorly and US images of the spinal cord can be acquired using a high-frequency linear abdominal transducer. Images of the normal spinal cord, normal variants (ventriculus terminalis, cyst of filum terminale) and spinal cord abnormalities (myelomeningocele, meningocele, diastematomyelia, tethered spinal cord and caudal regression syndrome) are presented. In this pictorial essay, comparison between images acquired with low- and high-frequency transducers are provided as well as correlation with postnatal data. In the normal spine, anatomical details such as the conus medullaris, the filum terminale and the nerve root bundles are exquisitely depicted, making it possible to differentiate normal variants from abnormalities. In abnormal cases, the position of the conus medullaris, its shape and the nerve roots can be analyzed in detail. We describe the benefits of using high-frequency linear transducers in US examination of the spinal cord, which is common after birth but has not been hitherto reported in fetuses.

Measures to define treatment response, such as no evidence of disease activity (NEDA), are routinely used in multiple sclerosis (MS) clinical practice. Although spinal cord involvement is a frequent feature of MS, its magnetic resonance imaging (MRI) monitoring is not routinely performed. To assess the impact of spinal cord MRI in the definition of NEDA in a cohort of people with MS (pwMS) with available spinal cord imaging performed as for routine monitoring. We included 115 pwMS undergoing treatment with first-line disease-modifying therapies (DMTs) and retrospectively analyzed the presence of NEDA in the whole cohort, either considering or not spinal cord imaging. When considering only clinical and brain MRI measures, 97 out of 115 pwMS (84.3%) satisfied the criteria for NEDA. In the same cohort, the number of pwMS with NEDA significantly decreased to 88 (76.5%) (p < 0.01) when considering also spinal cord imaging. These findings suggest that, in routine clinical practice, spinal cord MRI monitoring in pwMS under first-line DMTs leads to a slight but significant change in the proportion of subjects classified as clinically and radiologically stable according to the NEDA definition.