From Algorithm to Hardware: A Survey on Efficient and Safe Deployment of Deep Neural Networks.

Deep neural networks (DNNs) have been widely used in many artificial intelligence (AI) tasks. However, deploying them brings significant challenges due to the huge cost of memory, energy, and computation. To address these challenges, researchers have developed various model compression techniques such as model quantization and model pruning. Recently, there has been a surge in research on compression methods to achieve model efficiency while retaining performance. Furthermore, more and more works focus on customizing the DNN hardware accelerators to better leverage the model compression techniques. In addition to efficiency, preserving security and privacy is critical for deploying DNNs. However, the vast and diverse body of related works can be overwhelming. This inspires us to conduct a comprehensive survey on recent research toward the goal of high-performance, cost-efficient, and safe deployment of DNNs. Our survey first covers the mainstream model compression techniques, such as model quantization, model pruning, knowledge distillation, and optimizations of nonlinear operations. We then introduce recent advances in designing hardware accelerators that can adapt to efficient model compression approaches. In addition, we discuss how homomorphic encryption can be integrated to secure DNN deployment. Finally, we discuss several issues, such as hardware evaluation, generalization, and integration of various compression approaches. Overall, we aim to provide a big picture of efficient DNNs from algorithm to hardware accelerators and security perspectives.

Rigid spine injuries - A comprehensive review on diagnostic and therapeutic challenges.

Injuries to the rigid spine have a distinguished position in the broad spectrum of spinal injuries due to altered biomechanical properties. The rigid spine is more prone to fractures. Two ossification bone disorders that are of particular interest are Ankylosing Spondylitis (AS) and Diffuse Idiopathic Skeletal Hyperostosis (DISH). DISH is a non-inflammatory condition that leads to an anterolateral ossification of the spine. AS on the other hand is a chronic inflammatory disease that leads to cortical bone erosions and spinal ossifications. Both diseases gradually induce stiffening of the spine. The prevalence of DISH is age-related and is therefore higher in the older population. Although the prevalence of AS is not age-related the occurrence of spinal ossification is higher with increasing age. This association with age and the aging demographics in industrialized nations illustrate the need for medical professionals to be adequately informed and prepared. The aim of this narrating review is to give an overview on the diagnostic and therapeutic measures of the ankylosed spine. Because of highly unstable fracture configurations, injuries to the rigid spine are highly susceptible to neurological deficits. Diagnosing a fracture of the ankylosed spine on plain radiographs can be challenging. Moreover, since 8% of patients with ankylosing spine disorders (ASD) have multiple non-contagious fractures, a CT scan of the entire spine is highly recommended as the primary diagnostic tool. There are no consensus-based guidelines for the treatment of spinal fractures in ASD. The presence of neurological deficit or unstable fractures are absolute indications for surgical intervention. If conservative therapy is chosen, patients should be monitored closely to ensure that secondary neurologic deterioration does not occur. For the fractures that have to be treated surgically, stabilization of at least three segments above and below the fracture zone is recommended. These fractures mostly are treated via the posterior approach. Patients with AS or DISH share a significant risk for complications after a traumatic spine injury. The most frequent complications for patients with thoracolumbar burst fractures are respiratory failure, pseudoarthrosis, pneumonia, and implant failure.

Towards a standardized reporting of the impact of magnetic resonance imaging on the decision-making of thoracolumbar fractures without neurological deficit: Conceptual framework and proposed methodology.

Introduction: A recent meta-analysis showed that only four prior studies have shown that magnetic resonance imaging (MRI) can change the fracture classification in 17% and treatment decisions in 22% of cases. However, previous studies showed a wide methodological variability regarding the study population, the definition of posterior ligamentous complex (PLC) injury, and outcome measures. Research question: How can we standardize the reporting of the impact of MRI for neurologically intact patients with thoracolumbar fractures? Material and methods: All available literature regarding the impact of MRI on thoracolumbar fracture classification or decision-making were reviewed. Estimating the impact of MRI on the TLFs' classification is an exercise of analyzing the CTs' accuracy for PLC injury against MRI as a ''Gold standard''and should follow standardized checklists such as the Standards for the Reporting of Diagnostic Accuracy Studies. Additionally, specific issues related to TLFs should be addressed. Results: A standardized approach for reporting the impact of MRI in neurologically intact TLF patients was proposed. Regarding patient selection, restricting the inclusion of neurologically intact patients with A- and B-injuries is crucial. Image interpretation should be standardized regarding imaging protocol and appropriate criteria for PLC injury. The impact of MRI can be measured by either the rate of change in fracture classification or treatment decisions; the cons and pros of each measure is thoroughly discussed. Discussion and conclusion: We proposed a structured methodology for examining the impact of MRI on neurologically intact patients with TLFs, focusing on appropriate patient selection, standardizing image analysis, and clinically relevant outcome measures.

Traumatic thoracic spine fracture: can we predict when MRI would modify the fracture classification or decision-making compared to CT alone?

PURPOSE: To determine the impact of magnetic resonance imaging (MRI) on fracture classification for thoracic spine fractures (TSFs) compared to computed tomography (CT) alone. METHODS: This study was a retrospective review of 63 consecutive patients with TSFs who underwent CT and MRI within ten days of injury. Three reviewers classified all fractures according to the AOSpine Classification and the Thoracolumbar AOSpine Injury severity score (TLAOSIS). Posterior ligamentous complex (PLC) injury on MRI was defined by "black stripe discontinuity" and on CT by the presence of vertebral body translation, facet joint malalignment, horizontal laminar or spinous process fracture, and interspinous widening. The proportion of patients with AO type A/B/C and with TLAOSIS ≤ 5 and ≥ 6 was compared between CT and MRI. Classification and regression trees were used to create a series of predictive models for the probability of PLC injury in AO type A fractures. RESULTS: AO classification using CT was as follows: type A in 35 patients (55%), type B in 18 patients (29%), and type C in 10 patients (16%). Thirty-three patients (52%) had a TLAOSIS ≤5, while the remaining 30 (48%) had TLAOSI ≥6. The addition of MRI after CT upgraded type A to type B fractures in 10 patients (16%) and changed TL AOSIS from ≤5 to ≥6 in 8 cases (12.8%). Type A fractures with load sharing score (LSC) ≥6 had a 60% chance of upgrading to type B, while LSC <6 had a 12.5% chance of upgrading to type B. CONCLUSIONS: CT yielded (89%) accuracy in diagnosing PLC injury in TSFs. The addition of MRI after CT substantially changed the AO classification or TLAOISS, compared to CT alone, thus suggesting an added value of MRI for PLC assessment for TSFs classification.

How does vertical laminar fracture impact the decision-making in thoracolumbar fractures? A systematic scoping review and meta-analysis.

OBJECTIVE: Although vertical laminar fracture (VLF) is generally considered a severity marker for thoracolumbar fractures (TLFs), its exact role in decision-making has never been established. This scoping review aims to synthesize the research on VLF's role in the decision-making of TLFs. METHODS: A systematic review was conducted following PRISMA guidelines. We searched PubMed, Scopus, and Web of Science from inception to  June 11, 2023, for studies examining the association of VLF in thoracolumbar fractures with dural lacerations, neurological deficits, radiographic parameters, or treatment outcomes. Additionally, experimental studies that analyze the biomechanics of burst fractures with VLF were included. The studies extracted key findings, objectives, and patient population. A meta-analysis was performed for the association of VLF with dural laceration and neurological deficit, and ORs were pooled with a 95% confidence interval (CI). RESULTS: Twenty-eight studies were included in this systematic review, encompassing 2021 patients, and twelve were included in the meta-analysis. According to the main subject of the study, the association of VLF with a dural laceration (n = 14), neurological deficit (n = 4), radiographic parameters (n = 3), thoracolumbar fracture classification (n = 2), and treatment outcome (n = 2). Seven studies with a total of 1010 patients reported a significant association between VLF and neurological deficit (OR = 7.35, 95% CI [3.97, 14.25]; P < 0.001). The pooled OR estimates for VLF predicting dural lacerations were 7.75, 95% CI [2.41, 24.87]; P < 0.001). CONCLUSION: VLF may have several important diagnostic and therapeutic implications in managing TLFs. VLF may help to distinguish AO type A3 from A4 fractures. VLF may help to predict preoperatively the occurrence of dural laceration, thereby choosing the optimal surgical strategy. Clinical and biomechanical data suggest VLF may be a valuable modifier to guide the decision-making in burst fractures; however, more studies are needed to confirm its prognostic importance regarding treatment outcomes.

How frequently MRI modifies thoracolumbar fractures' classification or decision-making? A systematic review and meta-analysis.

PURPOSE: To provide the first meta-analysis of the impact of magnetic resonance imaging (MRI) on thoracolumbar fractures (TLFs) classification and decision-making. METHODS: A systematic review was conducted following PRISMA guidelines. We searched PubMed, Scopus, Cochrane, and Web of Science from inception to June 30, 2023 for studies evaluating the change in TLFs classification and treatment decisions after MRI. The studies extracted key findings, objectives, and patient population. A meta-analysis was performed for the pooled frequency of change in AO fracture classification or treatment decisions from surgical to conservative or vice versa after MRI. RESULTS: This meta-analysis included four studies comprising 554 patients. The pooled frequency of change in TLFs classification was 17% (95% CI 9-31%), and treatment decision was 22% (95% CI 11-40%). An upgrade from type A to type B was reported in 15.7% (95% CI 7.2-30.6%), and downgrading type B to type A in 1.2% (95% CI 0.17-8.3%). A change from conservative to surgery recommendation of 17% (95% CI 5.0-43%) was higher than a change from surgery to conservative 2% (95% CI 1-34%). CONCLUSIONS: MRI can significantly change the thoracolumbar classification and decision-making, primarily due to upgrading type A to type B fractures and changing from conservative to surgery, respectively. These findings suggest that MRI could change decision-making sufficiently to justify its use for TLFs. Type A subtypes, indeterminate PLC status, and spine regions might help to predict a change in TLFs' classification. However, more studies are needed to confirm the association of these variables with changes in treatment decisions to set the indications of MRI in neurologically intact patients with TLFs. An interactive version of our analysis can be accessed from here: https://databoard.shinyapps.io/mri_spine/ .

Using Equipoise to Determine the Radiographic Characteristics Leading to Agreement on Best Treatment for Thoracolumbar Burst Fractures Without Neurologic Deficits.

STUDY DESIGN: Retrospective analysis of prospectively collected data. OBJECTIVES: Our goal was to assess radiographic characteristics associated with agreement and disagreement in treatment recommendation in thoracolumbar (TL) burst fractures. METHODS: A panel of 22 AO Spine Knowledge Forum Trauma experts reviewed 183 cases and were asked to: (1) classify the fracture; (2) assess degree of certainty of PLC disruption; (3) assess degree of comminution; and (4) make a treatment recommendation. Equipoise threshold used was 77% (77:23 distribution of uncertainty or 17 vs 5 experts). Two groups were created: consensus vs equipoise. RESULTS: Of the 183 cases reviewed, the experts reached full consensus in only 8 cases (4.4%). Eighty-one cases (44.3%) were included in the agreement group and 102 cases (55.7%) in the equipoise group. A3/A4 fractures were more common in the equipoise group (92.0% vs 83.7%, P < .001). The agreement group had higher degree of certainty of PLC disruption [35.8% (SD 34.2) vs 27.6 (SD 27.3), P < .001] and more common use of the M1 modifier (44.3% vs 38.3%, P < .001). Overall, the degree of comminution was slightly higher in the equipoise group [47.8 (SD 20.5) vs 45.7 (SD 23.4), P < .001]. CONCLUSIONS: The agreement group had a higher degree of certainty of PLC injury and more common use of M1 modifier (more type B fractures). The equipoise group had more A3/A4 type fractures. Future studies are required to identify the role of comminution in decision making as degree of comminution was slightly higher in the equipoise group.

Interobserver Reliability in the Classification of Thoracolumbar Fractures Using the AO Spine TL Injury Classification System Among 22 Clinical Experts in Spine Trauma Care.

STUDY DESIGN: Reliability study utilizing 183 injury CT scans by 22 spine trauma experts with assessment of radiographic features, classification of injuries and treatment recommendations. OBJECTIVES: To assess the reliability of the AOSpine TL Injury Classification System (TLICS) including the categories within the classification and the M1 modifier. METHODS: Kappa and Intraclass correlation coefficients were produced. Associations of various imaging characteristics (comminution, PLC status) and treatment recommendations were analyzed through regression analysis. Multivariable logistic regression modeling was used for making predictive algorithms. RESULTS: Reliability of the AO Spine TLICS at differentiating A3 and A4 injuries (N = 71) (K = .466; 95% CI .458 - .474; P < .001) demonstrated moderate agreement. Similarly, the average intraclass correlation coefficient (ICC) amongst A3 and A4 injuries was excellent (ICC = .934; 95% CI .919 - .947; P < .001) and the ICC between individual measures was moderate (ICC = .403; 95% CI .351 - .461; P < .001). The overall agreement on the utilization of the M1 modifier amongst A3 and A4 injuries was fair (K = .161; 95% CI .151 - .171; P < .001). The ICC for PLC status in A3 and A4 injuries averaged across all measures was excellent (ICC = .936; 95% CI .922 - .949; P < .001). The M1 modifier suggests respondents are nearly 40% more confident that the PLC is injured amongst all injuries. The M1 modifier was employed at a higher frequency as injuries were classified higher in the classification system. CONCLUSIONS: The reliability of surgeons differentiating between A3 and A4 injuries in the AOSpine TLICS is substantial and the utilization of the M1 modifier occurs more frequently with higher grades in the system.

Expert Opinion, Real-World Classification, and Decision-Making in Thoracolumbar Burst Fractures Without Neurologic Deficits?

STUDY DESIGN: Retrospective analysis of prospectively collected data. OBJECTIVES: To compare decision-making between an expert panel and real-world spine surgeons in thoracolumbar burst fractures (TLBFs) without neurological deficits and analyze which factors influence surgical decision-making. METHODS: This study is a sub-analysis of a prospective observational study in TL fractures. Twenty two experts were asked to review 183 CT scans and recommend treatment for each fracture. The expert recommendation was based on radiographic review. RESULTS: Overall agreement between the expert panel and real-world surgeons regarding surgery was 63.2%. In 36.8% of cases, the expert panel recommended surgery that was not performed in real-world scenarios. Conversely, in cases where the expert panel recommended non-surgical treatment, only 38.6% received non-surgical treatment, while 61.4% underwent surgery. A separate analysis of A3 and A4 fractures revealed that expert panel recommended surgery for 30% of A3 injuries and 68% of A4 injuries. However, 61% of patients with both A3 and A4 fractures received surgery in the real world. Multivariate analysis demonstrated that a 1% increase in certainty of PLC injury led to a 4% increase in surgery recommendation among the expert panel, while a .2% increase in the likelihood of receiving surgery in the real world. CONCLUSION: Surgical decision-making varied between the expert panel and real-world treating surgeons. Differences appear to be less evident in A3/A4 burst fractures making this specific group of fractures a real challenge independent of the level of expertise.

The Influence of Comminution and Posterior Ligamentous Complex Integrity on Treatment Decision Making in Thoracolumbar Burst Fractures Without Neurologic Deficit?

STUDY DESIGN: A prospective study. OBJECTIVE: to evaluate the impact of vertebral body comminution and Posterior Ligamentous Complex (PLC) integrity on the treatment recommendations of thoracolumbar fractures among an expert panel of 22 spine surgeons. METHODS: A review of 183 prospectively collected thoracolumbar burst fracture computed tomography (CT) scans by an expert panel of 22 trauma spine surgeons to assess vertebral body comminution and PLC integrity. This study is a sub-study of a prospective observational study of thoracolumbar burst fractures (Spine TL A3/A4). Each expert was asked to grade the degree of comminution and certainty about the PLC disruption from 0 to 100, with 0 representing the intact vertebral body or intact PLC and 100 representing complete comminution or complete PLC disruption, respectively. RESULTS: ≥45% comminution had a 74% chance of having surgery recommended, while <25% comminution had an 86.3% chance of non-surgical treatment. A comminution from 25 to 45% had a 57% chance of non-surgical management. ≥55% PLC injury certainity had a 97% chance of having surgery, and ≥45-55% PLC injury certainty had a 65%. <20% PLC injury had a 64% chance of having non-operative treatment. A 20 to 45% PLC injury certainity had a 56% chance of non-surgical management. There was fair inter-rater agreement on the degree of comminution (ICC .57 [95% CI 0.52-.63]) and the PLC integrity (ICC .42 [95% CI 0.37-.48]). CONCLUSION: The study concludes that vetebral comminution and PLC integrity are major dterminant in decision making of thoracolumbar fractures without neurological deficit. However, more objective, reliable, and accurate methods of assessment of these variables are warranted.

The AO Spine Thoracolumbar Injury Classification System and Treatment Algorithm in Decision Making for Thoracolumbar Burst Fractures Without Neurologic Deficit.

STUDY DESIGN: Prospective Observational Study. OBJECTIVE: To determine the alignment of the AO Spine Thoracolumbar Injury Classification system and treatment algorithm with contemporary surgical decision making. METHODS: 183 cases of thoracolumbar burst fractures were reviewed by 22 AO Spine Knowledge Forum Trauma experts. These experienced clinicians classified the fracture morphology, integrity of the posterior ligamentous complex and degree of comminution. Management recommendations were collected. RESULTS: There was a statistically significant stepwise increase in rates of operative management with escalating category of injury (P < .001). An excellent correlation existed between recommended expert management and the actual treatment of each injury category: A0/A1/A2 (OR 1.09, 95% CI 0.70-1.69, P = .71), A3/4 (OR 1.62, 95% CI 0.98-2.66, P = .58) and B1/B2/C (1.00, 95% CI 0.87-1.14, P = .99). Thoracolumbar A4 fractures were more likely to be surgically stabilized than A3 fractures (68.2% vs 30.9%, P < .001). A modifier indicating indeterminate ligamentous injury increased the rate of operative management when comparing type B and C injuries to type A3/A4 injuries (OR 39.19, 95% CI 20.84-73.69, P < .01 vs OR 27.72, 95% CI 14.68-52.33, P < .01). CONCLUSIONS: The AO Spine Thoracolumbar Injury Classification system introduces fracture morphology in a rational and hierarchical manner of escalating severity. Thoracolumbar A4 complete burst fractures were more likely to be operatively managed than A3 fractures. Flexion-distraction type B injuries and translational type C injuries were much more likely to have surgery recommended than type A fractures regardless of the M1 modifier. A suspected posterior ligamentous injury increased the likelihood of surgeons favoring surgical stabilization.

Pediatric retroclival hematomas.

BACKGROUND: Traumatic retroclival hematomas (RCHs) are infrequent occurrences among the pediatric population. The existing body of research pertaining to these hematomas primarily consists of case reports or small case series, which do not provide adequate guidance for managing this condition. OBJECTIVE: This study aims to present a report on four cases of RCHs. Additionally, we aim to conduct a systematic review to consolidate the existing literature on pediatric RCHs. METHODS: The authors conducted a systematic review in accordance with the PRISMA and CARE guidelines. A multivariate logistic regression model was developed to evaluate the potential impact of various clinical variables on clinical outcomes. The study also documented four of our cases, one of which was a rare occurrence of spontaneous subdural RCH. RESULTS: A total of 62 traumatic RCHs have been documented in the literature. We documented three cases of traumatic RCHs and one case of spontaneous RCH. A systematic analysis of 65 traumatic RCHs was performed. Of trauma cases, 64.6% demonstrated craniocervical junction instability with 83.3% ligamentous involvement. Thirty-five patients were males. 50.7% were aged between 5 and 9 years. Cranial nerve palsies occurred in 29 patients (27 had abducent palsy), 26 of which resolved within 6 months of trauma. 23.5% underwent surgery, and 76.5% were conservatively managed. Surgeries targeted hematomas, hydrocephalus, or craniocervical instability. Approaches to hematomas included transclival and far/extreme lateral suboccipital approaches. Clinical outcome was good in 75.4% and intermediate or poor in 24.6%. Logistic regression suggested an association between craniocervical junction injuries and poor or intermediate outcomes (OR 4.88, 95% CI (1.17, 27.19), p = 0.04). CONCLUSION: Pediatric RCHs are mostly traumatic and extradural. Children between 5 and 9 years old are most vulnerable. Craniocervical junction injuries, mainly ligamentous, are common in RCHs and are associated with intermediate or poor outcomes. Cervical MRI could be important in cases of trauma to rule out ligamentous injuries of the craniocervical junction. The small size of RCHs should not exempt the careful assessment of craniocervical junction instability. Cranial nerve palsies are common and usually resolve within 6 months. Conservative treatment is typical unless brainstem compression, hydrocephalus, or craniocervical junction instability exists.

Which Morphological Features of Facet Diastasis Predict Thoracolumbar Posterior Ligamentous Complex Injury as Defined by Magnetic Resonance Imaging?

BACKGROUND: The association of various morphological features of facet diastasis with posterior ligamentous complex (PLC) injury has not been previously described. This study aims to determine the diagnostic value of facet diastasis subtypes for diagnosing thoracolumbar PLC injury. METHODS: We retrospectively reviewed 337 consecutive patients with acute thoracolumbar fractures who had computed tomography (CT) and magnetic resonance imaging (MRI) within 10 days of injury. Three and 5 reviewers evaluated MRI and CT images, respectively. Facet diastasis was subclassified as follows: Dislocated, no articular surface apposition; subluxed, incomplete articular surface apposition; and facet fracture articular process fractures which may be displaced ≥2 mm or otherwise undisplaced, facet joint widening (FJW) ≥ 3 mm. We examined the diagnostic accuracy and the multivariate associations of facet diastasis subtypes with PLC injury in MRI. RESULTS: Facet dislocation, subluxation, and displaced facet fracture yielded a high positive predictive value (PPV) for PLC injury (96%, 88%, and 94%, respectively). In contrast, undisplaced facet fracture and FJW yielded a moderate PPV for PLC injury (78%, and 45%, respectively). Facet dislocation, subluxation, and displaced facet fracture showed independent associations with PLC injury (adjusted odds ratio [AOR] = 38.4, 17.1, 13.4, respectively; P < 0.05). Undisplaced facet fracture and FJW were not associated with PLC injury (AOR = 3.9 [95% confidence interval, 0.49-38.4], P = 0.20) and (AOR = 1.94 [95% confidence interval, 0.48-7.13]; P = 0.20; P = 0.33), respectively. CONCLUSIONS: Facet dislocation, subluxation, and displaced facet fracture, but not undisplaced facet fracture or FJW, were independently associated with PLC injury. Therefore, we propose to define facet diastasis as a surrogate marker of PLC injury in MRI based on these morphologies.

A Proposal for a Standardized Imaging Algorithm to Improve the Accuracy and Reliability for the Diagnosis of Thoracolumbar Posterior Ligamentous Complex Injury in Computed Tomography and Magnetic Resonance Imaging.

STUDY DESIGN: Systematic Literature Review. OBJECTIVE: To propose a systematic imaging algorithm for diagnosing posterior ligamentous complex (PLC) injury in computed tomography (CT) and magnetic resonance imaging (MRI) to improve the reliability of PLC assessment. METHODS: A systematic review was conducted following PRISMA guidelines. The Scopus database was searched from its inception until July 21, 2022, for studies evaluating CT or MRI assessment of the PLC injury following thoracolumbar trauma. The studies extracted key findings, objectives, injury definitions, and radiographic modalities. RESULTS: Twenty-three studies were included in this systematic review, encompassing 2021 patients. Five studies evaluated the accuracy of MRI in detecting thoracolumbar PLC injury using intraoperative findings as a reference. These studies indicate that black stripe discontinuity due to supraspinous or ligamentum flavum rupture is a more specific criterion of PLC injury than high-signal intensity. Thirteen papers evaluated the accuracy or reliability of CT in detecting thoracolumbar PLC injury using MRI or intraoperative findings as a reference. The overall accuracy rate of CT in detecting PLC injury was 68-90%. Two studies evaluate the accuracy of combined CT findings, showing that ≥2 CT findings are associated with a positive predictive value of 88-91 %. Vertebral translation, facet joint malalignment, spinous process fracture, horizontal laminar fracture, and interspinous widening were independent predictors of PLC injury. CONCLUSION: We provided a comprehensive imaging algorithm for diagnosing PLC in CT and MRI based on available literature and our experience. The algorithm will potentially improve the accuracy and reliability of PLC assessment, however it needs multicentre prospective validation.