SARS-CoV-2 infection triggers profibrotic macrophage responses and lung fibrosis.

COVID-19-induced "acute respiratory distress syndrome" (ARDS) is associated with prolonged respiratory failure and high mortality, but the mechanistic basis of lung injury remains incompletely understood. Here, we analyze pulmonary immune responses and lung pathology in two cohorts of patients with COVID-19 ARDS using functional single-cell genomics, immunohistology, and electron microscopy. We describe an accumulation of CD163-expressing monocyte-derived macrophages that acquired a profibrotic transcriptional phenotype during COVID-19 ARDS. Gene set enrichment and computational data integration revealed a significant similarity between COVID-19-associated macrophages and profibrotic macrophage populations identified in idiopathic pulmonary fibrosis. COVID-19 ARDS was associated with clinical, radiographic, histopathological, and ultrastructural hallmarks of pulmonary fibrosis. Exposure of human monocytes to SARS-CoV-2, but not influenza A virus or viral RNA analogs, was sufficient to induce a similar profibrotic phenotype in vitro. In conclusion, we demonstrate that SARS-CoV-2 triggers profibrotic macrophage responses and pronounced fibroproliferative ARDS.

Clinically Applicable AI System for Accurate Diagnosis, Quantitative Measurements, and Prognosis of COVID-19 Pneumonia Using Computed Tomography.

Many COVID-19 patients infected by SARS-CoV-2 virus develop pneumonia (called novel coronavirus pneumonia, NCP) and rapidly progress to respiratory failure. However, rapid diagnosis and identification of high-risk patients for early intervention are challenging. Using a large computed tomography (CT) database from 3,777 patients, we developed an AI system that can diagnose NCP and differentiate it from other common pneumonia and normal controls. The AI system can assist radiologists and physicians in performing a quick diagnosis especially when the health system is overloaded. Significantly, our AI system identified important clinical markers that correlated with the NCP lesion properties. Together with the clinical data, our AI system was able to provide accurate clinical prognosis that can aid clinicians to consider appropriate early clinical management and allocate resources appropriately. We have made this AI system available globally to assist the clinicians to combat COVID-19.

Inactivation of CDK12 Delineates a Distinct Immunogenic Class of Advanced Prostate Cancer.

Using integrative genomic analysis of 360 metastatic castration-resistant prostate cancer (mCRPC) samples, we identified a novel subtype of prostate cancer typified by biallelic loss of CDK12 that is mutually exclusive with tumors driven by DNA repair deficiency, ETS fusions, and SPOP mutations. CDK12 loss is enriched in mCRPC relative to clinically localized disease and characterized by focal tandem duplications (FTDs) that lead to increased gene fusions and marked differential gene expression. FTDs associated with CDK12 loss result in highly recurrent gains at loci of genes involved in the cell cycle and DNA replication. CDK12 mutant cases are baseline diploid and do not exhibit DNA mutational signatures linked to defects in homologous recombination. CDK12 mutant cases are associated with elevated neoantigen burden ensuing from fusion-induced chimeric open reading frames and increased tumor T cell infiltration/clonal expansion. CDK12 inactivation thereby defines a distinct class of mCRPC that may benefit from immune checkpoint immunotherapy.

BACH2 immunodeficiency illustrates an association between super-enhancers and haploinsufficiency.

The transcriptional programs that guide lymphocyte differentiation depend on the precise expression and timing of transcription factors (TFs). The TF BACH2 is essential for T and B lymphocytes and is associated with an archetypal super-enhancer (SE). Single-nucleotide variants in the BACH2 locus are associated with several autoimmune diseases, but BACH2 mutations that cause Mendelian monogenic primary immunodeficiency have not previously been identified. Here we describe a syndrome of BACH2-related immunodeficiency and autoimmunity (BRIDA) that results from BACH2 haploinsufficiency. Affected subjects had lymphocyte-maturation defects that caused immunoglobulin deficiency and intestinal inflammation. The mutations disrupted protein stability by interfering with homodimerization or by causing aggregation. We observed analogous lymphocyte defects in Bach2-heterozygous mice. More generally, we observed that genes that cause monogenic haploinsufficient diseases were substantially enriched for TFs and SE architecture. These findings reveal a previously unrecognized feature of SE architecture in Mendelian diseases of immunity: heterozygous mutations in SE-regulated genes identified by whole-exome/genome sequencing may have greater significance than previously recognized.

Fossil evidence for a pharyngeal origin of the vertebrate pectoral girdle.

The origin of vertebrate paired appendages is one of the most investigated and debated examples of evolutionary novelty1-7. Paired appendages are widely considered as key innovations that enabled new opportunities for controlled swimming and gill ventilation and were prerequisites for the eventual transition from water to land. The past 150 years of debate8-10 has been shaped by two contentious theories4,5: the ventrolateral fin-fold hypothesis9,10 and the archipterygium hypothesis8. The latter proposes that fins and girdles evolved from an ancestral gill arch. Although studies in animal development have revived interest in this idea11-13, it is apparently unsupported by fossil evidence. Here we present palaeontological support for a pharyngeal basis for the vertebrate shoulder girdle. We use computed tomography scanning to reveal details of the braincase of Kolymaspis sibirica14, an Early Devonian placoderm fish from Siberia, that suggests a pharyngeal component of the shoulder. We combine these findings with refreshed comparative anatomy of placoderms and jawless outgroups to place the origin of the shoulder girdle on the sixth branchial arch. These findings provide a novel framework for understanding the origin of the pectoral girdle. Our evidence clarifies the location of the presumptive head-trunk interface in jawless fishes and explains the constraint on branchial arch number in gnathostomes15. The results revive a key aspect of the archipterygium hypothesis and help reconcile it with the ventrolateral fin-fold model.

Trial of Thrombectomy for Stroke with a Large Infarct of Unrestricted Size.

BACKGROUND: The use of thrombectomy in patients with acute stroke and a large infarct of unrestricted size has not been well studied. METHODS: We assigned, in a 1:1 ratio, patients with proximal cerebral vessel occlusion in the anterior circulation and a large infarct (as defined by an Alberta Stroke Program Early Computed Tomographic Score of ≤5; values range from 0 to 10) detected on magnetic resonance imaging or computed tomography within 6.5 hours after symptom onset to undergo endovascular thrombectomy and receive medical care (thrombectomy group) or to receive medical care alone (control group). The primary outcome was the score on the modified Rankin scale at 90 days (scores range from 0 to 6, with higher scores indicating greater disability). The primary safety outcome was death from any cause at 90 days, and an ancillary safety outcome was symptomatic intracerebral hemorrhage. RESULTS: A total of 333 patients were assigned to either the thrombectomy group (166 patients) or the control group (167 patients); 9 were excluded from the analysis because of consent withdrawal or legal reasons. The trial was stopped early because results of similar trials favored thrombectomy. Approximately 35% of the patients received thrombolysis therapy. The median modified Rankin scale score at 90 days was 4 in the thrombectomy group and 6 in the control group (generalized odds ratio, 1.63; 95% confidence interval [CI], 1.29 to 2.06; P<0.001). Death from any cause at 90 days occurred in 36.1% of the patients in the thrombectomy group and in 55.5% of those in the control group (adjusted relative risk, 0.65; 95% CI, 0.50 to 0.84), and the percentage of patients with symptomatic intracerebral hemorrhage was 9.6% and 5.7%, respectively (adjusted relative risk, 1.73; 95% CI, 0.78 to 4.68). Eleven procedure-related complications occurred in the thrombectomy group. CONCLUSIONS: In patients with acute stroke and a large infarct of unrestricted size, thrombectomy plus medical care resulted in better functional outcomes and lower mortality than medical care alone but led to a higher incidence of symptomatic intracerebral hemorrhage. (Funded by Montpellier University Hospital; LASTE ClinicalTrials.gov number, NCT03811769.).

A new coronavirus associated with human respiratory disease in China.

Emerging infectious diseases, such as severe acute respiratory syndrome (SARS) and Zika virus disease, present a major threat to public health1-3. Despite intense research efforts, how, when and where new diseases appear are still a source of considerable uncertainty. A severe respiratory disease was recently reported in Wuhan, Hubei province, China. As of 25 January 2020, at least 1,975 cases had been reported since the first patient was hospitalized on 12 December 2019. Epidemiological investigations have suggested that the outbreak was associated with a seafood market in Wuhan. Here we study a single patient who was a worker at the market and who was admitted to the Central Hospital of Wuhan on 26 December 2019 while experiencing a severe respiratory syndrome that included fever, dizziness and a cough. Metagenomic RNA sequencing4 of a sample of bronchoalveolar lavage fluid from the patient identified a new RNA virus strain from the family Coronaviridae, which is designated here 'WH-Human 1' coronavirus (and has also been referred to as '2019-nCoV'). Phylogenetic analysis of the complete viral genome (29,903 nucleotides) revealed that the virus was most closely related (89.1% nucleotide similarity) to a group of SARS-like coronaviruses (genus Betacoronavirus, subgenus Sarbecovirus) that had previously been found in bats in China5. This outbreak highlights the ongoing ability of viral spill-over from animals to cause severe disease in humans.

Phase-contrast virtual chest radiography.

Respiratory X-ray imaging enhanced by phase contrast has shown improved airway visualization in animal models. Limitations in current X-ray technology have nevertheless hindered clinical translation, leaving the potential clinical impact an open question. Here, we explore phase-contrast chest radiography in a realistic in silico framework. Specifically, we use preprocessed virtual patients to generate in silico chest radiographs by Fresnel-diffraction simulations of X-ray wave propagation. Following a reader study conducted with clinical radiologists, we predict that phase-contrast edge enhancement will have a negligible impact on improving solitary pulmonary nodule detection (6 to 20 mm). However, edge enhancement of bronchial walls visualizes small airways (< 2 mm), which are invisible in conventional radiography. Our results show that phase-contrast chest radiography could play a future role in observing small-airway obstruction (e.g., relevant for asthma or early-stage chronic obstructive pulmonary disease), which cannot be directly visualized using current clinical methods, thereby motivating the experimental development needed for clinical translation. Finally, we discuss quantitative requirements on distances and X-ray source/detector specifications for clinical implementation of phase-contrast chest radiography.

CT or Invasive Coronary Angiography in Stable Chest Pain.

BACKGROUND: In the diagnosis of obstructive coronary artery disease (CAD), computed tomography (CT) is an accurate, noninvasive alternative to invasive coronary angiography (ICA). However, the comparative effectiveness of CT and ICA in the management of CAD to reduce the frequency of major adverse cardiovascular events is uncertain. METHODS: We conducted a pragmatic, randomized trial comparing CT with ICA as initial diagnostic imaging strategies for guiding the treatment of patients with stable chest pain who had an intermediate pretest probability of obstructive CAD and were referred for ICA at one of 26 European centers. The primary outcome was major adverse cardiovascular events (cardiovascular death, nonfatal myocardial infarction, or nonfatal stroke) over 3.5 years. Key secondary outcomes were procedure-related complications and angina pectoris. RESULTS: Among 3561 patients (56.2% of whom were women), follow-up was complete for 3523 (98.9%). Major adverse cardiovascular events occurred in 38 of 1808 patients (2.1%) in the CT group and in 52 of 1753 (3.0%) in the ICA group (hazard ratio, 0.70; 95% confidence interval [CI], 0.46 to 1.07; P = 0.10). Major procedure-related complications occurred in 9 patients (0.5%) in the CT group and in 33 (1.9%) in the ICA group (hazard ratio, 0.26; 95% CI, 0.13 to 0.55). Angina during the final 4 weeks of follow-up was reported in 8.8% of the patients in the CT group and in 7.5% of those in the ICA group (odds ratio, 1.17; 95% CI, 0.92 to 1.48). CONCLUSIONS: Among patients referred for ICA because of stable chest pain and intermediate pretest probability of CAD, the risk of major adverse cardiovascular events was similar in the CT group and the ICA group. The frequency of major procedure-related complications was lower with an initial CT strategy. (Funded by the European Union Seventh Framework Program and others; DISCHARGE ClinicalTrials.gov number, NCT02400229.).

VascuViz: a multimodality and multiscale imaging and visualization pipeline for vascular systems biology.

Despite advances in imaging, image-based vascular systems biology has remained challenging because blood vessel data are often available only from a single modality or at a given spatial scale, and cross-modality data are difficult to integrate. Therefore, there is an exigent need for a multimodality pipeline that enables ex vivo vascular imaging with magnetic resonance imaging, computed tomography and optical microscopy of the same sample, while permitting imaging with complementary contrast mechanisms from the whole-organ to endothelial cell spatial scales. To achieve this, we developed 'VascuViz'-an easy-to-use method for simultaneous three-dimensional imaging and visualization of the vascular microenvironment using magnetic resonance imaging, computed tomography and optical microscopy in the same intact, unsectioned tissue. The VascuViz workflow permits multimodal imaging with a single labeling step using commercial reagents and is compatible with diverse tissue types and protocols. VascuViz's interdisciplinary utility in conjunction with new data visualization approaches opens up new vistas in image-based vascular systems biology.

NIEND: neuronal image enhancement through noise disentanglement.

MOTIVATION: The full automation of digital neuronal reconstruction from light microscopic images has long been impeded by noisy neuronal images. Previous endeavors to improve image quality can hardly get a good compromise between robustness and computational efficiency. RESULTS: We present the image enhancement pipeline named Neuronal Image Enhancement through Noise Disentanglement (NIEND). Through extensive benchmarking on 863 mouse neuronal images with manually annotated gold standards, NIEND achieves remarkable improvements in image quality such as signal-background contrast (40-fold) and background uniformity (10-fold), compared to raw images. Furthermore, automatic reconstructions on NIEND-enhanced images have shown significant improvements compared to both raw images and images enhanced using other methods. Specifically, the average F1 score of NIEND-enhanced reconstructions is 0.88, surpassing the original 0.78 and the second-ranking method, which achieved 0.84. Up to 52% of reconstructions from NIEND-enhanced images outperform all other four methods in F1 scores. In addition, NIEND requires only 1.6 s on average for processing 256 × 256 × 256-sized images, and images after NIEND attain a substantial average compression rate of 1% by LZMA. NIEND improves image quality and neuron reconstruction, providing potential for significant advancements in automated neuron morphology reconstruction of petascale. AVAILABILITY AND IMPLEMENTATION: The study is conducted based on Vaa3D and Python 3.10. Vaa3D is available on GitHub (https://github.com/Vaa3D). The proposed NIEND method is implemented in Python, and hosted on GitHub along with the testing code and data (https://github.com/zzhmark/NIEND). The raw neuronal images of mouse brains can be found at the BICCN's Brain Image Library (BIL) (https://www.brainimagelibrary.org). The detailed list and associated meta information are summarized in Supplementary Table S3.

Impact of radiological response and pattern of progression in patients with HCC treated by atezolizumab-bevacizumab.

BACKGROUND AND AIMS: We aim to assess the role of radiological response to atezolizumab-bevacizumab in patients with HCC to predict overall survival. APPROACH AND RESULTS: We retrospectively included patients with HCC treated by atezolizumab-bevacizumab in 2 tertiary centers. A retrospective blinded analysis was performed by 2 radiologists to assess Response Evaluation Criteria in Solid Tumor (RECIST 1.1) and modified RECIST (mRECIST) criteria at 12 weeks. Imaging response and treatment decisions in the multidisciplinary tumor board at 12 weeks were registered. Among 125 patients, 9.6% and 20.8% had a response, 39.2% and 35.2% had stable disease, and 51.2% and 44% had progression, according to RECIST 1.1 and mRECIST, respectively, with a substantial interobserver agreement (k coefficient=0.79). Metastasis was independently associated with a higher risk of progression. Patients classified as responders did not reach median survival, which was 16.2 and 15.9 months for patients classified as stable and 9.1 and 9.0 months for patients classified as progressors, in RECIST 1.1 and mRECIST criteria, respectively. We observed a wide variability in the identification of progression in the multidisciplinary tumor board in clinical practice compared with the blind evaluation by radiologists mainly due to discrepancy in the evaluation of the increase in size of intrahepatic lesions. The appearance of new extrahepatic lesions or vascular invasion lesions was associated with a worse overall survival ( p =0.032). CONCLUSIONS: RECIST 1.1 and mRECIST criteria predict overall survival with more responders identified by mRECIST and the appearance of new extrahepatic lesion or vascular invasion was associated with a poor prognosis. A noticeable discrepancy was observed between patients classified as progressors at reviewing and the decision reached during the multidisciplinary tumor board.

CT radiomics analysis discriminates pulmonary lesions in patients with pulmonary MALT lymphoma and non-pulmonary MALT lymphoma.

PURPOSE: The aim of this study is to create and validate a radiomics model based on CT scans, enabling the distinction between pulmonary mucosa-associated lymphoid tissue (MALT) lymphoma and other pulmonary lesion causes. METHODS: Patients diagnosed with primary pulmonary MALT lymphoma and lung infections at Fuzhou Pulmonary Hospital were randomly assigned to either a training group or a validation group. Meanwhile, individuals diagnosed with primary pulmonary MALT lymphoma and lung infections at Fujian Provincial Cancer Hospital were chosen as the external test group. We employed ITK-SNAP software for delineating the Region of Interest (ROI) within the images. Subsequently, we extracted radiomics features and convolutional neural networks using PyRadiomics, a component of the Onekey AI software suite. Relevant radiomic features were selected to build an intelligent diagnostic prediction model utilizing CT images, and the model's efficacy was assessed in both the validation group and the external test group. RESULTS: Leveraging radiomics, ten distinct features were carefully chosen for analysis. Subsequently, this study employed the machine learning techniques of Logistic Regression (LR), Support Vector Machine (SVM), and k-Nearest Neighbors (KNN) to construct models using these ten selected radiomics features within the training groups. Among these, SVM exhibited the highest performance, achieving an accuracy of 0.868, 0.870, and 0.90 on the training, validation, and external testing groups, respectively. For LR, the accuracy was 0.837, 0.863, and 0.90 on the training, validation, and external testing groups, respectively. For KNN, the accuracy was 0.884, 0.859, and 0.790 on the training, validation, and external testing groups, respectively. CONCLUSION: We established a noninvasive radiomics model utilizing CT imaging to diagnose pulmonary MALT lymphoma associated with pulmonary lesions. This model presents a promising adjunct tool to enhance diagnostic specificity for pulmonary MALT lymphoma, particularly in populations where pulmonary lesion changes may be attributed to other causes.

MRUNet-3D: A multi-stride residual 3D UNet for lung nodule segmentation.

Obtaining an accurate segmentation of the pulmonary nodules in computed tomography (CT) images is challenging. This is due to: (1) the heterogeneous nature of the lung nodules; (2) comparable visual characteristics between the nodules and their surroundings. A robust multi-scale feature extraction mechanism that can effectively obtain multi-scale representations at a granular level can improve segmentation accuracy. As the most commonly used network in lung nodule segmentation, UNet, its variants, and other image segmentation methods lack this robust feature extraction mechanism. In this study, we propose a multi-stride residual 3D UNet (MRUNet-3D) to improve the segmentation accuracy of lung nodules in CT images. It incorporates a multi-slide Res2Net block (MSR), which replaces the simple sequence of convolution layers in each encoder stage to effectively extract multi-scale features at a granular level from different receptive fields and resolutions while conserving the strengths of 3D UNet. The proposed method has been extensively evaluated on the publicly available LUNA16 dataset. Experimental results show that it achieves competitive segmentation performance with an average dice similarity coefficient of 83.47 % and an average surface distance of 0.35 mm on the dataset. More notably, our method has proven to be robust to the heterogeneity of lung nodules. It has also proven to perform better at segmenting small lung nodules. Ablation studies have shown that the proposed MSR and RFIA modules are fundamental to improving the performance of the proposed model.

MSO-GP: 3-D segmentation of large and complex conjoined tree structures.

Robust segmentation of large and complex conjoined tree structures in 3-D is a major challenge in computer vision. This is particularly true in computational biology, where we often encounter large data structures in size, but few in number, which poses a hard problem for learning algorithms. We show that merging multiscale opening with geodesic path propagation, can shed new light on this classic machine vision challenge, while circumventing the learning issue by developing an unsupervised visual geometry approach (digital topology/morphometry). The novelty of the proposed MSO-GP method comes from the geodesic path propagation being guided by a skeletonization of the conjoined structure that helps to achieve robust segmentation results in a particularly challenging task in this area, that of artery-vein separation from non-contrast pulmonary computed tomography angiograms. This is an important first step in measuring vascular geometry to then diagnose pulmonary diseases and to develop image-based phenotypes. We first present proof-of-concept results on synthetic data, and then verify the performance on pig lung and human lung data with less segmentation time and user intervention needs than those of the competing methods.

Interactions between vascular burden and amyloid-ß pathology on trajectories of tau accumulation.

Cerebrovascular pathology often co-exists with Alzheimer's disease pathology and can contribute to Alzheimer's disease-related clinical progression. However, the degree to which vascular burden contributes to Alzheimer's disease pathological progression is still unclear. This study aimed to investigate interactions between vascular burden and amyloid-ß pathology on both baseline tau tangle load and longitudinal tau accumulation. We included 1229 participants from the Swedish BioFINDER-2 Study, including cognitively unimpaired and impaired participants with and without biomarker-confirmed amyloid-ß pathology. All underwent baseline tau-PET (18F-RO948), and a subset (n = 677) underwent longitudinal tau-PET after 2.5 ± 1.0 years. Tau-PET uptake was computed for a temporal meta-region-of-interest. We focused on four main vascular imaging features and risk factors: microbleeds; white matter lesion volume; stroke-related events (infarcts, lacunes and haemorrhages); and the Framingham Heart Study Cardiovascular Disease risk score. To validate our in vivo results, we examined 1610 autopsy cases from an Arizona-based neuropathology cohort on three main vascular pathological features: cerebral amyloid angiopathy; white matter rarefaction; and infarcts. For the in vivo cohort, primary analyses included age-, sex- and APOE ɛ4-corrected linear mixed models between tau-PET (outcome) and interactions between time, amyloid-ß and each vascular feature (predictors). For the neuropathology cohort, age-, sex- and APOE ɛ4-corrected linear models between tau tangle density (outcome) and an interaction between plaque density and each vascular feature (predictors) were performed. In cognitively unimpaired individuals, we observed a significant interaction between microbleeds and amyloid-ß pathology on greater baseline tau load (ß = 0.68, P < 0.001) and longitudinal tau accumulation (ß = 0.11, P < 0.001). For white matter lesion volume, we did not observe a significant independent interaction effect with amyloid-ß on tau after accounting for microbleeds. In cognitively unimpaired individuals, we further found that stroke-related events showed a significant negative interaction with amyloid-ß on longitudinal tau (ß = -0.08, P < 0.001). In cognitively impaired individuals, there were no significant interaction effects between cerebrovascular and amyloid-ß pathology at all. In the neuropathology dataset, the in vivo observed interaction effects between cerebral amyloid angiopathy and plaque density (ß = 0.38, P < 0.001) and between infarcts and plaque density (ß = -0.11, P = 0.005) on tau tangle density were replicated. To conclude, we demonstrated that cerebrovascular pathology-in the presence of amyloid-ß pathology-modifies tau accumulation in early stages of Alzheimer's disease. More specifically, the co-occurrence of microbleeds and amyloid-ß pathology was associated with greater accumulation of tau aggregates during early disease stages. This opens the possibility that interventions targeting microbleeds may attenuate the rate of tau accumulation in Alzheimer's disease.

Serum biomarkers and disease progression in CT-negative mild traumatic brain injury.

Blood proteins are emerging as potential biomarkers for mild traumatic brain injury (mTBI). Molecular pathology of mTBI underscores the critical roles of neuronal injury, neuroinflammation, and vascular health in disease progression. However, the temporal profile of blood biomarkers associated with the aforementioned molecular pathology after CT-negative mTBI, their diagnostic and prognostic potential, and their utility in monitoring white matter integrity and progressive brain atrophy remain unclear. Thus, we investigated serum biomarkers and neuroimaging in a longitudinal cohort, including 103 CT-negative mTBI patients and 66 matched healthy controls (HCs). Angiogenic biomarker vascular endothelial growth factor (VEGF) exhibited the highest area under the curve of 0.88 in identifying patients from HCs. Inflammatory biomarker interleukin-1ß and neuronal cell body injury biomarker ubiquitin carboxyl-terminal hydrolase L1 were elevated in acute-stage patients and associated with deterioration of cognitive function from acute-stage to 6-12 mo post-injury period. Notably, axonal injury biomarker neurofilament light (NfL) was elevated in acute-stage patients, with higher levels associated with impaired white matter integrity in acute-stage and progressive gray and white matter atrophy from 3- to 6-12 mo post-injury period. Collectively, our findings emphasized the potential clinical value of serum biomarkers, particularly NfL and VEGF, in diagnosing mTBI and monitoring disease progression.

Predicting Obstructive Sleep Apnea Based on Computed Tomography Scans Using Deep Learning Models.

Rationale: The incidence of clinically undiagnosed obstructive sleep apnea (OSA) is high among the general population because of limited access to polysomnography. Computed tomography (CT) of craniofacial regions obtained for other purposes can be beneficial in predicting OSA and its severity. Objectives: To predict OSA and its severity based on paranasal CT using a three-dimensional deep learning algorithm. Methods: One internal dataset (N = 798) and two external datasets (N = 135 and N = 85) were used in this study. In the internal dataset, 92 normal participants and 159 with mild, 201 with moderate, and 346 with severe OSA were enrolled to derive the deep learning model. A multimodal deep learning model was elicited from the connection between a three-dimensional convolutional neural network-based part treating unstructured data (CT images) and a multilayer perceptron-based part treating structured data (age, sex, and body mass index) to predict OSA and its severity. Measurements and Main Results: In a four-class classification for predicting the severity of OSA, the AirwayNet-MM-H model (multimodal model with airway-highlighting preprocessing algorithm) showed an average accuracy of 87.6% (95% confidence interval [CI], 86.8-88.6%) in the internal dataset and 84.0% (95% CI, 83.0-85.1%) and 86.3% (95% CI, 85.3-87.3%) in the two external datasets, respectively. In the two-class classification for predicting significant OSA (moderate to severe OSA), the area under the receiver operating characteristic curve, accuracy, sensitivity, specificity, and F1 score were 0.910 (95% CI, 0.899-0.922), 91.0% (95% CI, 90.1-91.9%), 89.9% (95% CI, 88.8-90.9%), 93.5% (95% CI, 92.7-94.3%), and 93.2% (95% CI, 92.5-93.9%), respectively, in the internal dataset. Furthermore, the diagnostic performance of the Airway Net-MM-H model outperformed that of the other six state-of-the-art deep learning models in terms of accuracy for both four- and two-class classifications and area under the receiver operating characteristic curve for two-class classification (P < 0.001). Conclusions: A novel deep learning model, including a multimodal deep learning model and an airway-highlighting preprocessing algorithm from CT images obtained for other purposes, can provide significantly precise outcomes for OSA diagnosis.