Automating Scoliosis Measurements in Radiographic Studies with Machine Learning: Comparing Artificial Intelligence and Clinical Reports.

Scoliosis is a condition of abnormal lateral spinal curvature affecting an estimated 2 to 3% of the US population, or seven million people. The Cobb angle is the standard measurement of spinal curvature in scoliosis but is known to have high interobserver and intraobserver variability. Thus, the objective of this study was to build and validate a system for automatic quantitative evaluation of the Cobb angle and to compare AI generated and human reports in the clinical setting. After IRB was obtained, we retrospectively collected 2150 frontal view scoliosis radiographs at a tertiary referral center (January 1, 2019, to January 1, 2021, ≥ 16 years old, no hardware). The dataset was partitioned into 1505 train (70%), 215 validation (10%), and 430 test images (20%). All thoracic and lumbar vertebral bodies were segmented with bounding boxes, generating approximately 36,550 object annotations that were used to train a Faster R-CNN Resnet-101 object detection model. A controller algorithm was written to localize vertebral centroid coordinates and derive the Cobb properties (angle and endplate) of dominant and secondary curves. AI-derived Cobb angle measurements were compared to the clinical report measurements, and the Spearman rank-order demonstrated significant correlation (0.89, p < 0.001). Mean difference between AI and clinical report angle measurements was 7.34° (95% CI: 5.90-8.78°), which is similar to published literature (up to 10°). We demonstrate the feasibility of an AI system to automate measurement of level-by-level spinal angulation with performance comparable to radiologists.

Deep Learning for Automated Classification of Inferior Vena Cava Filter Types on Radiographs.

PURPOSE: To demonstrate the feasibility and evaluate the performance of a deep-learning convolutional neural network (CNN) classification model for automated identification of different types of inferior vena cava (IVC) filters on radiographs. MATERIALS AND METHODS: In total, 1,375 cropped radiographic images of 14 types of IVC filters were collected from patients enrolled in a single-center IVC filter registry, with 139 images withheld as a test set and the remainder used to train and validate the classification model. Image brightness, contrast, intensity, and rotation were varied to augment the training set. A 50-layer ResNet architecture with fixed pre-trained weights was trained using a soft margin loss over 50 epochs. The final model was evaluated on the test set. RESULTS: The CNN classification model achieved a F1 score of 0.97 (0.92-0.99) for the test set overall and of 1.00 for 10 of 14 individual filter types. Of the 139 test set images, 4 (2.9%) were misidentified, all mistaken for other filter types that appear highly similar. Heat maps elucidated salient features for each filter type that the model used for class prediction. CONCLUSIONS: A CNN classification model was successfully developed to identify 14 types of IVC filters on radiographs and demonstrated high performance. Further refinement and testing of the model is necessary before potential real-world application.

MRI UTE-T2* shows high incidence of cartilage subsurface matrix changes 2 years after ACL reconstruction.

Alteration of deep cartilage matrix has been observed following anterior cruciate ligament (ACL) injury, evidenced by elevated MRI UTE-T2* values measured in small, 2-D cartilage regions of interest. This Level I diagnostic study seeks to more thoroughly evaluate deep cartilage matrix changes to medial tibiofemoral UTE-T2* maps 2 years after ACL reconstruction and examine the relative utilities of 3-D compared to 2-D assessments of cartilage UTE-T2* maps. Thirty-eight ACL-reconstructed and 20 uninjured subjects underwent MRI UTE-T2* mapping. "Small" single mid-sagittal 2-D and larger 3-D "tread mark" regions of interest were manually segmented and found to be correlated in medial cartilage (r > 0.58, p < 0.005). 3-D analyses of UTE-T2* maps showed differences to medial tibial cartilage between ACL-reconstructed and uninjured subjects (p = 0.007) that were not detected by smaller 2-D regions (p > 0.46). Quantitative comparisons show 14/38 (37%) ACL-reconstructed subjects have values >2 standard deviations higher than uninjured controls. Among a subset of ACL-reconstructed subjects with no morphologic MRI evidence of medial tibiofemoral cartilage or meniscal pathology (n = 12), elevated UTE-T2* values in "small" 2-D femoral (p = 0.011), but not larger 3-D tread mark regions of interest (p > 0.13), were observed. These data show the utility of 2-D UTE-T2* assessments of mid-sagittal weight-bearing regions of medial femoral cartilage for identifying subclinical deep cartilage matrix changes 2 years after ACLR. Clinical Significance: Mid-sagittal single slice 2-D UTE-T2* mapping may be an efficient means to assess medial femoral cartilage for subsurface matrix changes early after ACL reconstruction while 3-D assessments provide additional sensitivity to changes in the medial tibial plateau. © 2018 Orthopaedic Research Society. Published by Wiley Periodicals, Inc. J Orthop Res 37:370-377, 2019.

Bone Tumor Diagnosis Using a Naïve Bayesian Model of Demographic and Radiographic Features.

Because many bone tumors have a variety of appearances and are uncommon, few radiologists develop sufficient expertise to guide optimal management. Bayesian inference can guide decision-making by computing probabilities of multiple diagnoses to generate a differential. We built and validated a naïve Bayes machine (NBM) that processes 18 demographic and radiographic features. We reviewed over 1664 analog radiographic cases of bone tumors and selected 811 cases (66 diagnoses) for annotation using a quantitative imaging platform. Leave-one-out cross validation was performed. Primary accuracy was defined as the correct pathological diagnosis as the top machine prediction. Differential accuracy was defined as whether the correct pathological diagnosis was within the top three predictions. For the 29 most common diagnoses (710 cases), primary accuracy was 44%, and differential accuracy was 60%. For the top 10 most common diagnoses (478 cases), primary accuracy was 62%, and differential accuracy was 80%. The machine returned relevant diagnoses for the majority of unknown test cases and may be a feasible alternative to machine learning approaches such as deep neural networks or support vector machines that typically require larger training data (our model required a minimum of five samples per diagnosis) and are "black boxes" (our model can provide details of probability calculations to identify features that most significantly contribute to truth diagnoses). Finally, our Bayes model was designed to scale and "learn" from external data, enabling incorporation of outside knowledge such as Dahlin's Bone Tumors, a reference of anatomic and demographic statistics of more than 10,000 tumors.

MR imaging near metallic implants using MAVRIC SL: initial clinical experience at 3T.

RATIONALE AND OBJECTIVES: To compare the effectiveness of multiacquisition with variable resonance image combination selective (MAVRIC SL) with conventional two-dimensional fast spin-echo (2D-FSE) magnetic resonance (MR) techniques at 3T in imaging patients with a variety of metallic implants. MATERIALS AND METHODS: Twenty-one 3T MR studies were obtained in 19 patients with different types of metal implants. Paired MAVRIC SL and 2D-FSE sequences were reviewed by two radiologists and compared for in-plane and through-plane metal artifact, visualization of the bone implant interface and surrounding soft tissues, blurring, and overall image quality using a two-tailed Wilcoxon signed rank test. The area of artifact on paired images was measured and compared using a paired Wilcoxon signed rank test. Changes in patient management resulting from MAVRIC SL imaging were documented. RESULTS: Significantly less in-plane and through-plane artifact was seen with MAVRIC SL, with improved visualization of the bone-implant interface and surrounding soft tissues, and superior overall image quality (P = .0001). Increased blurring was seen with MAVRIC SL (P = .0016). MAVRIC SL significantly decreased the image artifact compared to 2D-FSE (P = .0001). Inclusion of MAVRIC SL to the imaging protocol determined the need for surgery or type of surgery in five patients and ruled out the need for surgery in 13 patients. In three patients, the area of interest was well seen on both MAVRIC SL and 2D-FSE images, so the addition of MAVRIC had no effect on patient management. CONCLUSIONS: Imaging around metal implants with MAVRIC SL at 3T significantly improved image quality and decreased image artifact compared to conventional 2D-FSE imaging techniques and directly impacted patient management.

Quantitative Magnetic Resonance Imaging UTE-T2* Mapping of Cartilage and Meniscus Healing After Anatomic Anterior Cruciate Ligament Reconstruction.

BACKGROUND: An anterior cruciate ligament (ACL) injury greatly increases the risk for premature knee osteoarthritis (OA). Improved diagnosis and staging of early disease are needed to develop strategies to delay or prevent disabling OA. PURPOSE: Novel magnetic resonance imaging (MRI) ultrashort echo time (UTE)-T2(*) mapping was evaluated against clinical metrics of cartilage health in cross-sectional and longitudinal studies of human participants before and after ACL reconstruction (ACLR) to show reversible deep subsurface cartilage and meniscus matrix changes. STUDY DESIGN: Cohort study (diagnosis/prognosis); Level of evidence, 2. METHODS: Forty-two participants (31 undergoing anatomic ACLR; 11 uninjured) underwent 3-T MRI inclusive of a sequence capturing short and ultrashort T2 signals. An arthroscopic examination of the medial meniscus was performed, and modified Outerbridge grades were assigned to the central and posterior medial femoral condyle (cMFC and pMFC, respectively) of ACL-reconstructed patients. Two years after ACLR, 16 patients underwent the same 3-T MRI. UTE-T2(*) maps were generated for the posterior medial meniscus (pMM), cMFC, pMFC, and medial tibial plateau (MTP). Cross-sectional evaluations of UTE-T2(*) and arthroscopic data along with longitudinal analyses of UTE-T2(*) changes were performed. RESULTS: Arthroscopic grades showed that 74% (23/31) of ACL-reconstructed patients had intact cMFC cartilage (Outerbridge grade 0 and 1) and that 90% (28/31) were Outerbridge grade 0 to 2. UTE-T2(*) values in deep cMFC and pMFC cartilage varied significantly with injury status and arthroscopic grade (Outerbridge grade 0-2: n = 39; P = .03 and .04, respectively). Pairwise comparisons showed UTE-T2(*) differences between uninjured controls (n = 11) and patients with arthroscopic Outerbridge grade 0 for the cMFC (n = 12; P = .01) and arthroscopic Outerbridge grade 1 for the pMFC (n = 11; P = .01) only and not individually between arthroscopic Outerbridge grade 0, 1, and 2 of ACL-reconstructed patients (P > .05). Before ACLR, UTE-T2(*) values of deep cMFC and pMFC cartilage of ACL-reconstructed patients were a respective 43% and 46% higher than those of uninjured controls (14.1 ± 5.5 vs 9.9 ± 2.3 milliseconds [cMFC] and 17.4 ± 7.0 vs 11.9 ± 2.4 milliseconds [pMFC], respectively; P = .02 for both). In longitudinal analyses, preoperative elevations in UTE-T2(*) values in deep pMFC cartilage and the pMM in those with clinically intact menisci decreased to levels similar to those in uninjured controls (P = .02 and .005, respectively), suggestive of healing. No decrease in UTE-T2(*) values for the MFC and new elevation in UTE-T2(*) values for the submeniscus MTP were observed in those with meniscus tears. CONCLUSION: This study shows that novel UTE-T2(*) mapping demonstrates changes in cartilage deep tissue health according to joint injury status as well as a potential for articular cartilage and menisci to heal deep tissue injuries. Further clinical studies of UTE-T2(*) mapping are needed to determine if it can be used to identify joints at risk for rapid degeneration and to monitor effects of new treatments to delay or prevent the development of OA.

Automatic retrieval of bone fracture knowledge using natural language processing.

Natural language processing (NLP) techniques to extract data from unstructured text into formal computer representations are valuable for creating robust, scalable methods to mine data in medical documents and radiology reports. As voice recognition (VR) becomes more prevalent in radiology practice, there is opportunity for implementing NLP in real time for decision-support applications such as context-aware information retrieval. For example, as the radiologist dictates a report, an NLP algorithm can extract concepts from the text and retrieve relevant classification or diagnosis criteria or calculate disease probability. NLP can work in parallel with VR to potentially facilitate evidence-based reporting (for example, automatically retrieving the Bosniak classification when the radiologist describes a kidney cyst). For these reasons, we developed and validated an NLP system which extracts fracture and anatomy concepts from unstructured text and retrieves relevant bone fracture knowledge. We implement our NLP in an HTML5 web application to demonstrate a proof-of-concept feedback NLP system which retrieves bone fracture knowledge in real time.

Pattern of 18F-FDG uptake in the spinal cord in patients with non-central nervous system malignancy.

STUDY DESIGN: Retrospective review. OBJECTIVE: To (1) propose a standard method to quantitate 2-deoxy-2-[18F]-fluoro-D-glucose (18F-FDG) uptake in the spinal cord and (2) use this methodology to retrospectively characterize the pattern of uptake within the entire spinal cord using whole-body positron emission tomography/computed tomography (PET/CT) imaging. SUMMARY OF BACKGROUND DATA: A physiologic understanding of glucose metabolism within the spinal cord may provide insight regarding infectious, inflammatory, vascular, and neoplastic spinal cord diseases. METHODS: Institutional review board approval was obtained. A total of 131 consecutive whole-body PET/CT studies from July to August 2004 were reviewed, and using exclusionary criteria of: (1) severe spinal arthropathy or curvature, (2) motion artifact, (3) canal hardware, (4) spinal tumor, and (5) marrow hyperplasia, 92 studies of neurologically intact patients (49 men and 43 women) were selected for a retrospective review of spinal cord 18F-FDG activity. The transaxial CT was used to define the canal and circular regions of interests were placed within the canal at the level of the vertebral body midpoint from C1 to L3. Region of interest total count, area, and maximum standardized uptake value (SUVmax) were recorded. Measurements at L5 served as an internal control. For comparative analysis, the cord-to-background (CTB) ratio was defined as spinal cord SUVmax to L5 SUVmax. RESULTS: Mean CTB decreased along each spinal level from cranial to caudal (P < 0.001). Significant relative increases were observed at the T11-T12 vertebral body levels (P < 0.001). Although insignificant, a relative increase was observed at C4. No significant interactions of age or sex on CTB were observed. CONCLUSION: The pattern of 18F-FDG uptake within the spinal cord, observed in patients with non-central nervous system malignancy, may be helpful in understanding glucose physiology of spinal cord diseases and warrants further research.

Evaluation of negation and uncertainty detection and its impact on precision and recall in search.

Radiology reports contain information that can be mined using a search engine for teaching, research, and quality assurance purposes. Current search engines look for exact matches to the search term, but they do not differentiate between reports in which the search term appears in a positive context (i.e., being present) from those in which the search term appears in the context of negation and uncertainty. We describe RadReportMiner, a context-aware search engine, and compare its retrieval performance with a generic search engine, Google Desktop. We created a corpus of 464 radiology reports which described at least one of five findings (appendicitis, hydronephrosis, fracture, optic neuritis, and pneumonia). Each report was classified by a radiologist as positive (finding described to be present) or negative (finding described to be absent or uncertain). The same reports were then classified by RadReportMiner and Google Desktop. RadReportMiner achieved a higher precision (81%), compared with Google Desktop (27%; p < 0.0001). RadReportMiner had a lower recall (72%) compared with Google Desktop (87%; p = 0.006). We conclude that adding negation and uncertainty identification to a word-based radiology report search engine improves the precision of search results over a search engine that does not take this information into account. Our approach may be useful to adopt into current report retrieval systems to help radiologists to more accurately search for radiology reports.

Informatics in radiology: RADTF: a semantic search-enabled, natural language processor-generated radiology teaching file.

Storing and retrieving radiology cases is an important activity for education and clinical research, but this process can be time-consuming. In the process of structuring reports and images into organized teaching files, incidental pathologic conditions not pertinent to the primary teaching point can be omitted, as when a user saves images of an aortic dissection case but disregards the incidental osteoid osteoma. An alternate strategy for identifying teaching cases is text search of reports in radiology information systems (RIS), but retrieved reports are unstructured, teaching-related content is not highlighted, and patient identifying information is not removed. Furthermore, searching unstructured reports requires sophisticated retrieval methods to achieve useful results. An open-source, RadLex(®)-compatible teaching file solution called RADTF, which uses natural language processing (NLP) methods to process radiology reports, was developed to create a searchable teaching resource from the RIS and the picture archiving and communication system (PACS). The NLP system extracts and de-identifies teaching-relevant statements from full reports to generate a stand-alone database, thus converting existing RIS archives into an on-demand source of teaching material. Using RADTF, the authors generated a semantic search-enabled, Web-based radiology archive containing over 700,000 cases with millions of images. RADTF combines a compact representation of the teaching-relevant content in radiology reports and a versatile search engine with the scale of the entire RIS-PACS collection of case material.

Conservation of the RB1 gene in human and primates.

Mutations in the RB1 gene are associated with retinoblastoma, which has served as an important model for understanding hereditary predisposition to cancer. Despite the great scrutiny that RB1 has enjoyed as the prototypical tumor suppressor gene, it has never been the object of a comprehensive survey of sequence variation in diverse human populations and primates. Therefore, we analyzed the coding (2,787 bp) and adjacent intronic and untranslated (7,313 bp) sequences of RB1 in 137 individuals from a wide range of ethnicities, including 19 Asian Indian hereditary retinoblastoma cases, and five primate species. Aside from nine apparently disease-associated mutations, 52 variants were identified. They included six singleton, coding variants that comprised five amino acid replacements and one silent site. Nucleotide diversity of the coding region (pi=0.0763+/-1.35 x 10(-4)) was 52 times lower than that of the noncoding regions (pi=3.93+/-5.26 x 10(-4)), indicative of significant sequence conservation. The occurrence of purifying selection was corroborated by phylogeny-based maximum likelihood analysis of the RB1 sequences of human and five primates, which yielded an estimated ratio of replacement to silent substitutions (omega) of 0.095 across all lineages. RB1 displayed extensive linkage disequilibrium over 174 kb, and only four unique recombination events, two in Africa and one each in Europe and Southwest Asia, were observed. Using a parsimony approach, 15 haplotypes could be inferred. Ten were found in Africa, though only 12.4% of the 274 chromosomes screened were of African origin. In non-Africans, a single haplotype accounted for from 63 to 84% of all chromosomes, most likely the consequence of natural selection and a significant bottleneck in effective population size during the colonization of the non-African continents.