A deep learning-based technique for the diagnosis of epidural spinal cord compression on thoracolumbar CT.

PURPOSE: To develop a deep learning (DL) model for epidural spinal cord compression (ESCC) on CT, which will aid earlier ESCC diagnosis for less experienced clinicians. METHODS: We retrospectively collected CT and MRI data from adult patients with suspected ESCC at a tertiary referral institute from 2007 till 2020. A total of 183 patients were used for training/validation of the DL model. A separate test set of 40 patients was used for DL model evaluation and comprised 60 staging CT and matched MRI scans performed with an interval of up to 2 months. DL model performance was compared to eight readers: one musculoskeletal radiologist, two body radiologists, one spine surgeon, and four trainee spine surgeons. Diagnostic performance was evaluated using inter-rater agreement, sensitivity, specificity and AUC. RESULTS: Overall, 3115 axial CT slices were assessed. The DL model showed high kappa of 0.872 for normal, low and high-grade ESCC (trichotomous), which was superior compared to a body radiologist (R4, κ = 0.667) and all four trainee spine surgeons (κ range = 0.625-0.838)(all p < 0.001). In addition, for dichotomous normal versus any grade of ESCC detection, the DL model showed high kappa (κ = 0.879), sensitivity (91.82), specificity (92.01) and AUC (0.919), with the latter AUC superior to all readers (AUC range = 0.732-0.859, all p < 0.001). CONCLUSION: A deep learning model for the objective assessment of ESCC on CT had comparable or superior performance to radiologists and spine surgeons. Earlier diagnosis of ESCC on CT could reduce treatment delays, which are associated with poor outcomes, increased costs, and reduced survival.

Improved Productivity Using Deep Learning-assisted Reporting for Lumbar Spine MRI.

Background Lumbar spine MRI studies are widely used for back pain assessment. Interpretation involves grading lumbar spinal stenosis, which is repetitive and time consuming. Deep learning (DL) could provide faster and more consistent interpretation. Purpose To assess the speed and interobserver agreement of radiologists for reporting lumbar spinal stenosis with and without DL assistance. Materials and Methods In this retrospective study, a DL model designed to assist radiologists in the interpretation of spinal canal, lateral recess, and neural foraminal stenoses on lumbar spine MRI scans was used. Randomly selected lumbar spine MRI studies obtained in patients with back pain who were 18 years and older over a 3-year period, from September 2015 to September 2018, were included in an internal test data set. Studies with instrumentation and scoliosis were excluded. Eight radiologists, each with 2-13 years of experience in spine MRI interpretation, reviewed studies with and without DL model assistance with a 1-month washout period. Time to diagnosis (in seconds) and interobserver agreement (using Gwet κ) were assessed for stenosis grading for each radiologist with and without the DL model and compared with test data set labels provided by an external musculoskeletal radiologist (with 32 years of experience) as the reference standard. Results Overall, 444 images in 25 patients (mean age, 51 years ± 20 [SD]; 14 women) were evaluated in a test data set. DL-assisted radiologists had a reduced interpretation time per spine MRI study, from a mean of 124-274 seconds (SD, 25-88 seconds) to 47-71 seconds (SD, 24-29 seconds) (P < .001). DL-assisted radiologists had either superior or equivalent interobserver agreement for all stenosis gradings compared with unassisted radiologists. DL-assisted general and in-training radiologists improved their interobserver agreement for four-class neural foraminal stenosis, with κ values of 0.71 and 0.70 (with DL) versus 0.39 and 0.39 (without DL), respectively (both P < .001). Conclusion Radiologists who were assisted by deep learning for interpretation of lumbar spinal stenosis on MRI scans showed a marked reduction in reporting time and superior or equivalent interobserver agreement for all stenosis gradings compared with radiologists who were unassisted by deep learning. © RSNA, 2022 Online supplemental material is available for this article. See also the editorial by Hayashi in this issue.

Impact of contrast-enhanced mammography in surgical management of breast cancers for women with dense breasts: a dual-center, multi-disciplinary study in Asia.

OBJECTIVE: To evaluate the impact of pre-operative contrast-enhanced mammography (CEM) in breast cancer patients with dense breasts. METHODS: We conducted a retrospective review of 232 histologically proven breast cancers in 200 women (mean age: 53.4 years ± 10.2) who underwent pre-surgical CEM imaging across two Asian institutions (Singapore and Taiwan). Majority (95.5%) of patients had dense breast tissue (BI-RADS category C or D). Surgical decision was recorded in a simulated blinded multi-disciplinary team setting on two separate scenarios: (i) pre-CEM setting with standard imaging, and clinical and histopathological results; and (ii) post-CEM setting with new imaging and corresponding histological findings from CEM. Alterations in surgical plan (if any) because of CEM imaging were recorded. Predictors CEM of patients who benefitted from surgical plan alterations were evaluated using logistic regression. RESULTS: CEM resulted in altered surgical plans in 36 (18%) of 200 patients in this study. CEM discovered clinically significant larger tumor size or extent in 24 (12%) patients and additional tumors in 12 (6%) patients. CEM also detected additional benign/false-positive lesions in 13 (6.5%) of the 200 patients. Significant predictors of patients who benefitted from surgical alterations found on multivariate analysis were pre-CEM surgical decision for upfront breast conservation (OR, 7.7; 95% CI, 1.9-32.1; p = 0.005), architectural distortion on mammograms (OR, 7.6; 95% CI, 1.3-42.9; p = .022), and tumor size of ≥ 1.5 cm (OR, 1.5; 95% CI, 1.0-2.2; p = .034). CONCLUSION: CEM is an effective imaging technique for pre-surgical planning for Asian breast cancer patients with dense breasts. KEY POINTS: • CEM significantly altered surgical plans in 18% (nearly 1 in 5) of this Asian study cohort with dense breasts. • Significant patient and imaging predictors for surgical plan alteration include (i) patients considered for upfront breast-conserving surgery; (ii) architectural distortion lesions; and (iii) tumor size of ≥ 1.5 cm. • Additional false-positive/benign lesions detected through CEM were uncommon, affecting only 6.5% of the study cohort.

Effect of Training Data Volume on Performance of Convolutional Neural Network Pneumothorax Classifiers.

Large datasets with high-quality labels required to train deep neural networks are challenging to obtain in the radiology domain. This work investigates the effect of training dataset size on the performance of deep learning classifiers, focusing on chest radiograph pneumothorax detection as a proxy visual task in the radiology domain. Two open-source datasets (ChestX-ray14 and CheXpert) comprising 291,454 images were merged and convolutional neural networks trained with stepwise increase in training dataset sizes. Model iterations at each dataset volume were evaluated on an external test set of 525 emergency department chest radiographs. Learning curve analysis was performed to fit the observed AUCs for all models generated. For all three network architectures tested, model AUCs and accuracy increased rapidly from 2 × 103 to 20 × 103 training samples, with more gradual increase until the maximum training dataset size of 291 × 103 images. AUCs for models trained with the maximum tested dataset size of 291 × 103 images were significantly higher than models trained with 20 × 103 images: ResNet-50: AUC20k = 0.86, AUC291k = 0.95, p < 0.001; DenseNet-121 AUC20k = 0.85, AUC291k = 0.93, p < 0.001; EfficientNet AUC20k = 0.92, AUC 291 k = 0.98, p < 0.001. Our study established learning curves describing the relationship between dataset training size and model performance of deep learning convolutional neural networks applied to a typical radiology binary classification task. These curves suggest a point of diminishing performance returns for increasing training data volumes, which algorithm developers should consider given the high costs of obtaining and labelling radiology data.

Deep Learning Model for Automated Detection and Classification of Central Canal, Lateral Recess, and Neural Foraminal Stenosis at Lumbar Spine MRI.

Background Assessment of lumbar spinal stenosis at MRI is repetitive and time consuming. Deep learning (DL) could improve -productivity and the consistency of reporting. Purpose To develop a DL model for automated detection and classification of lumbar central canal, lateral recess, and neural -foraminal stenosis. Materials and Methods In this retrospective study, lumbar spine MRI scans obtained from September 2015 to September 2018 were included. Studies of patients with spinal instrumentation or studies with suboptimal image quality, as well as postgadolinium studies and studies of patients with scoliosis, were excluded. Axial T2-weighted and sagittal T1-weighted images were used. Studies were split into an internal training set (80%), validation set (9%), and test set (11%). Training data were labeled by four radiologists using predefined gradings (normal, mild, moderate, and severe). A two-component DL model was developed. First, a convolutional neural network (CNN) was trained to detect the region of interest (ROI), with a second CNN for classification. An internal test set was labeled by a musculoskeletal radiologist with 31 years of experience (reference standard) and two subspecialist radiologists (radiologist 1: A.M., 5 years of experience; radiologist 2: J.T.P.D.H., 9 years of experience). DL model performance on an external test set was evaluated. Detection recall (in percentage), interrater agreement (Gwet κ), sensitivity, and specificity were calculated. Results Overall, 446 MRI lumbar spine studies were analyzed (446 patients; mean age ± standard deviation, 52 years ± 19; 240 women), with 396 patients in the training (80%) and validation (9%) sets and 50 (11%) in the internal test set. For internal testing, DL model and radiologist central canal recall were greater than 99%, with reduced neural foramina recall for the DL model (84.5%) and radiologist 1 (83.9%) compared with radiologist 2 (97.1%) (P < .001). For internal testing, dichotomous classification (normal or mild vs moderate or severe) showed almost-perfect agreement for both radiologists and the DL model, with respective κ values of 0.98, 0.98, and 0.96 for the central canal; 0.92, 0.95, and 0.92 for lateral recesses; and 0.94, 0.95, and 0.89 for neural foramina (P < .001). External testing with 100 MRI scans of lumbar spines showed almost perfect agreement for the DL model for dichotomous classification of all ROIs (κ, 0.95-0.96; P < .001). Conclusion A deep learning model showed comparable agreement with subspecialist radiologists for detection and classification of central canal and lateral recess stenosis, with slightly lower agreement for neural foraminal stenosis at lumbar spine MRI. © RSNA, 2021 Online supplemental material is available for this article. See also the editorial by Hayashi in this issue.

Architecture distortion score (ADS) in malignancy risk stratification of architecture distortion on contrast-enhanced digital mammography.

OBJECTIVE: To develop a risk predictor model in evaluation of tomosynthesis-detected architectural distortion (AD) based on characteristics of contrast-enhanced digital mammography (CEDM). METHODS: Ninety-four AD lesions on CEDM in combination with tomosynthesis were retrospectively reviewed from 92 consecutive women (mean age, 52.4 years ± 7.9) with abnormal diagnostic or screening mammography. CEDM results were correlated with histology of ADs using cross-tabulation for statistical analysis. Predictors for risk of malignancy from CEDM characteristics (background parenchyma enhancement, degree of AD enhancement, enhancing morphology, size of enhancement, and enhancing spiculations) and patient's age were evaluated using logistic regression. We propose a sum score, termed AD score (ADS), for risk stratification and corresponding suggested BI-RADS category. RESULTS: Thirty-three of ninety-four (35.1%) of detected AD lesions were malignant. The sensitivity, specificity, PPV, and NPV of CEDM in evaluation of malignant AD are 100%, 42.6%, 48.5%, and 100%, respectively. Absence of AD enhancement on CEDM is highly indicative of no underlying malignancy. On multivariate analysis, the predictors on CEDM with statistical significance are (1) marked intensity of AD enhancement (OR, 22.6; 95%CI 3.1, 166.6; p = .002); and (2) presence of enhancing spiculations (OR, 9.1; 95%CI 2.2, 36.5; p = .002). A prediction model whose scores (ADS) given by ranking of OR of all predictors with AUC of 0.934 and Brier score of 0.0956 was developed. CONCLUSION: ADS-based lesion characterization on CEDM enables risk assessment of tomosynthesis-detected AD lesions. KEY POINTS: • Architecture distortions presenting with marked enhancement intensity and presence of enhancing spiculations are highly associated with risk of malignancy. • Absence of architecture distortion enhancement in minimal or mild background parenchyma enhancement on CEDM indicates low risk of breast malignancy (NPV = 100%).

Large-scale genotyping identifies a new locus at 22q13.2 associated with female breast size.

BACKGROUND: Individual differences in breast size are a conspicuous feature of variation in human females and have been associated with fecundity and advantage in selection of mates. To identify common variants that are associated with breast size, we conducted a large-scale genotyping association meta-analysis in 7169 women of European descent across three independent sample collections with digital or screen film mammograms. METHODS: The samples consisted of the Swedish KARMA, LIBRO-1 and SASBAC studies genotyped on iCOGS, a custom illumina iSelect genotyping array comprising of 211 155 single nucleotide polymorphisms (SNPs) designed for replication and fine mapping of common and rare variants with relevance to breast, ovary and prostate cancer. Breast size of each subject was ascertained by measuring total breast area (mm(2)) on a mammogram. RESULTS: We confirm genome-wide significant associations at 8p11.23 (rs10086016, p=1.3×10(-14)) and report a new locus at 22q13 (rs5995871, p=3.2×10(-8)). The latter region contains the MKL1 gene, which has been shown to impact endogenous oestrogen receptor α transcriptional activity and is recruited on oestradiol sensitive genes. We also replicated previous genome-wide association study findings for breast size at four other loci. CONCLUSIONS: A new locus at 22q13 may be associated with female breast size.

Intra-articular and Peri-articular Tumours and Tumour Mimics- What a Clinician and Onco-imaging Radiologist Should Know.

Definitive determination of the cause of articular swelling may be difficult based on just the clinical symptoms, physical examinations and laboratory tests. Joint disorders fall under the realms of rheumatology and general orthopaedics; however, patients with joint conditions manifesting primarily as intra-articular and peri-articular soft tissue swelling may at times be referred to an orthopaedic oncology department with suspicion of a tumour. In such a situation, an onco-radiologist needs to think beyond the usual neoplastic lesions and consider the diagnoses of various non-neoplastic arthritic conditions that may be clinically masquerading as masses. Differential diagnoses of articular lesions include infectious and non-infectious synovial proliferative processes, degenerative lesions, deposition diseases, vascular malformations, benign and malignant neoplasms and additional miscellaneous conditions. Many of these diseases have specific imaging findings. Knowledge of these radiological characteristics in an appropriate clinical context will allow for a more confident diagnosis.

Local cOinage and Hospital Equipment Index (Lo Hei): projectile distance of Singapore coinage and healthcare-related equipment in a 3T magnetic resonance imaging scanner.

INTRODUCTION: Modern magnetic resonance imaging (MRI) scanners utilise superconducting magnets that are permanently active. Patients and healthcare professionals have been known to unintentionally introduce ferromagnetic objects into the scanning room. In this study, we evaluated the projectile risk of Singapore coinage as well as some common healthcare equipment within a 3 T MRI scanner. METHODS: A rig termed 'Object eNtry Guidance and Linear Acceleration Instrument' (ONG LAI) was custom-built to facilitate safe trajectory of the putative ferromagnetic objects. A ballistic gel target was utilised as a human tissue surrogate to estimate tissue penetration. The point at which objects would self-propel towards the scanner was named 'Huge Unintended Acceleration Towards Actual Harm (HUAT AH)'. RESULTS: Singapore third-series coins (10-cent to 1-dollar coins) are highly ferromagnetic and would accelerate towards the MRI scanner from more than one metre away. Cannulas with their needles are ferromagnetic and would self-propel towards the scanner from a distance of 20 cm. Standard surgical masks are ferromagnetic and may lose their sealing efficacy when they are worn too close to the magnet. Among the tested objects, a can of pineapple drink (Lee Pineapple Juice) had the highest HUAT AH at a distance of more than 1.5 m. CONCLUSION: Some local coinage and commonly found objects within a healthcare setting demonstrate ferromagnetic activity with projectile potential from a distance of more than 1 m. Patients and healthcare professionals should be cognisant of the risk associated with introducing these objects into the MRI scanning room.

Application of Machine Learning for Differentiating Bone Malignancy on Imaging: A Systematic Review.

An accurate diagnosis of bone tumours on imaging is crucial for appropriate and successful treatment. The advent of Artificial intelligence (AI) and machine learning methods to characterize and assess bone tumours on various imaging modalities may assist in the diagnostic workflow. The purpose of this review article is to summarise the most recent evidence for AI techniques using imaging for differentiating benign from malignant lesions, the characterization of various malignant bone lesions, and their potential clinical application. A systematic search through electronic databases (PubMed, MEDLINE, Web of Science, and clinicaltrials.gov) was conducted according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. A total of 34 articles were retrieved from the databases and the key findings were compiled and summarised. A total of 34 articles reported the use of AI techniques to distinguish between benign vs. malignant bone lesions, of which 12 (35.3%) focused on radiographs, 12 (35.3%) on MRI, 5 (14.7%) on CT and 5 (14.7%) on PET/CT. The overall reported accuracy, sensitivity, and specificity of AI in distinguishing between benign vs. malignant bone lesions ranges from 0.44-0.99, 0.63-1.00, and 0.73-0.96, respectively, with AUCs of 0.73-0.96. In conclusion, the use of AI to discriminate bone lesions on imaging has achieved a relatively good performance in various imaging modalities, with high sensitivity, specificity, and accuracy for distinguishing between benign vs. malignant lesions in several cohort studies. However, further research is necessary to test the clinical performance of these algorithms before they can be facilitated and integrated into routine clinical practice.

Common Healthcare Related Instruments Subjected To Magnetic Attraction Study (CHRISTMAS): prospective in situ experimental study.

OBJECTIVE: To investigate the behaviour of common healthcare related objects in a 3 tesla (T) MRI (magnetic resonance imaging) scanner, examining their ability to self-propel towards the scanner bore and their potential for tissue penetration. DESIGN: Prospective in situ experimental study. SETTING: Clinical 3 T MRI scanner. Customised rig designed and built to guide objects towards the scanner bore. PARTICIPANTS: 12 categories of objects commonly found in hospitals, or on patients or healthcare professionals, or near an MRI scanning room. Human tissue penetration simulated with ballistic gel (Federal Bureau of Investigation and North Atlantic Treaty Organisation graded). MAIN OUTCOME MEASURES: SANTA (site where applied newtonian mechanics triggers acceleration) measurements and depth of tissue penetration of the objects. RESULTS: SANTA measurements ranged from 0 cm for the 20 pence, 50 pence, and £2 coins to 152-161 cm for a knife and the biscuit tins. One penny, two pence, five pence, and 10 pence coins showed self-propulsion and acceleration towards the scanner bore at a distance >100 cm from the gantry entry point. Linear regression analysis showed no apparent correlation between the weight of the objects and their SANTA measurements (R2<0.1). Only five objects penetrated the ballistic gel (simulated human tissue). The deepest penetration was by the knife (5.5 cm), closely followed by the teaspoon (5.0 cm), fork (4.0 cm), spoon (3.5 cm), and a 10 pence coin (0.5 cm). Although the biscuit tins did not penetrate the simulated human tissue, they exerted substantial impact force which could potentially cause bone fractures. A smartphone, digital thermometer, metallic credit card, and pen torch remained fully functional after several passes into the MRI scanner. No discernible loss of image quality for the MRI scanner after the experiments was found. CONCLUSIONS: The study highlights the potential for harm (major tissue damage and bone fractures) when commonly found objects in a healthcare setting are unintentionally brought into the MRI scanner room. Patients and healthcare professionals need to be aware of the dangers associated with bringing ferromagnetic objects into the MRI environment.

SIR Flat CAP (Safety In Radiology - Flat-Packed Compact Airborne Precaution): A Low-Cost, Portable, Negative-Pressure Isolation Barrier Shield for Protecting Frontline Healthcare Workers.

Introduction Multiple barrier shields have been described since the start of the COVID-19 pandemic. Most of these are bulky and designed for use in the main anesthetic or radiology departments. We developed a portable, negative-pressure barrier shield designed specifically for portable ultrasound examinations. A novel supine cough generation model was developed together with a reverse qualitative fit test to simulate real-world aerosol droplet generation and dispersion for evaluating the effectiveness of the barrier shield. We report the technical specifications of this design, named "SIR Flat CAP" from Safety In Radiology - Flat-packed Compact Airborne Precaution, as well as its performance in reducing the spread of droplets and aerosols.  Methods The barrier shield was constructed using 1 mm acrylic panels, clear packing tape, foam double-sided tape, and surgical drapes. Negative pressure was provided via hospital wall suction. A supine cough generation model was developed to simulate cough droplet dispersal. A reverse qualitative fit test was used to assess for airborne transmission of microdroplets. Results The supine cough generation model was able to replicate similar results to previously reported supine human cough generation dispersion. The use of the barrier shield with negative-pressure suction prevented the escape of visible droplets, and no airborne microdroplets were detected by reverse qualitative fit testing from the containment area. Conclusions The barrier shield significantly reduces the escape of visible and airborne droplets from the containment area, providing an additional layer of protection to front-line sonographers.

Deep learning assessment compared to radiologist reporting for metastatic spinal cord compression on CT.

Introduction: Metastatic spinal cord compression (MSCC) is a disastrous complication of advanced malignancy. A deep learning (DL) algorithm for MSCC classification on CT could expedite timely diagnosis. In this study, we externally test a DL algorithm for MSCC classification on CT and compare with radiologist assessment. Methods: Retrospective collection of CT and corresponding MRI from patients with suspected MSCC was conducted from September 2007 to September 2020. Exclusion criteria were scans with instrumentation, no intravenous contrast, motion artefacts and non-thoracic coverage. Internal CT dataset split was 84% for training/validation and 16% for testing. An external test set was also utilised. Internal training/validation sets were labelled by radiologists with spine imaging specialization (6 and 11-years post-board certification) and were used to further develop a DL algorithm for MSCC classification. The spine imaging specialist (11-years expertise) labelled the test sets (reference standard). For evaluation of DL algorithm performance, internal and external test data were independently reviewed by four radiologists: two spine specialists (Rad1 and Rad2, 7 and 5-years post-board certification, respectively) and two oncological imaging specialists (Rad3 and Rad4, 3 and 5-years post-board certification, respectively). DL model performance was also compared against the CT report issued by the radiologist in a real clinical setting. Inter-rater agreement (Gwet's kappa) and sensitivity/specificity/AUCs were calculated. Results: Overall, 420 CT scans were evaluated (225 patients, mean age=60 ± 11.9[SD]); 354(84%) CTs for training/validation and 66(16%) CTs for internal testing. The DL algorithm showed high inter-rater agreement for three-class MSCC grading with kappas of 0.872 (p<0.001) and 0.844 (p<0.001) on internal and external testing, respectively. On internal testing DL algorithm inter-rater agreement (κ=0.872) was superior to Rad 2 (κ=0.795) and Rad 3 (κ=0.724) (both p<0.001). DL algorithm kappa of 0.844 on external testing was superior to Rad 3 (κ=0.721) (p<0.001). CT report classification of high-grade MSCC disease was poor with only slight inter-rater agreement (κ=0.027) and low sensitivity (44.0), relative to the DL algorithm with almost-perfect inter-rater agreement (κ=0.813) and high sensitivity (94.0) (p<0.001). Conclusion: Deep learning algorithm for metastatic spinal cord compression on CT showed superior performance to the CT report issued by experienced radiologists and could aid earlier diagnosis.

State-of-the-Art Imaging Techniques in Metastatic Spinal Cord Compression.

Metastatic Spinal Cord Compression (MSCC) is a debilitating complication in oncology patients. This narrative review discusses the strengths and limitations of various imaging modalities in diagnosing MSCC, the role of imaging in stereotactic body radiotherapy (SBRT) for MSCC treatment, and recent advances in deep learning (DL) tools for MSCC diagnosis. PubMed and Google Scholar databases were searched using targeted keywords. Studies were reviewed in consensus among the co-authors for their suitability before inclusion. MRI is the gold standard of imaging to diagnose MSCC with reported sensitivity and specificity of 93% and 97% respectively. CT Myelogram appears to have comparable sensitivity and specificity to contrast-enhanced MRI. Conventional CT has a lower diagnostic accuracy than MRI in MSCC diagnosis, but is helpful in emergent situations with limited access to MRI. Metal artifact reduction techniques for MRI and CT are continually being researched for patients with spinal implants. Imaging is crucial for SBRT treatment planning and three-dimensional positional verification of the treatment isocentre prior to SBRT delivery. Structural and functional MRI may be helpful in post-treatment surveillance. DL tools may improve detection of vertebral metastasis and reduce time to MSCC diagnosis. This enables earlier institution of definitive therapy for better outcomes.