Deep semi-supervised learning for automatic segmentation of inferior alveolar nerve using a convolutional neural network

Background The inferior alveolar nerve (IAN) innervates and regulates the sensation of the mandibular teeth and lower lip. The position of the IAN should be monitored prior to surgery. Therefore, a study using artificial intelligence (AI) was planned to image and track the position of the IAN automatically for a quicker and safer surgery. Methods A total of 138 cone-beam computed tomography datasets (Internal: 98, External: 40) collected from multiple centers (three hospitals) were used in the study. A customized 3D nnU-Net was used for image segmentation. Active learning, which consists of three steps, was carried out in iterations for 83 datasets with cumulative additions after each step. Subsequently, the accuracy of the model for IAN segmentation was evaluated using the 50 datasets. The accuracy by deriving the dice similarity coefficient (DSC) value and the segmentation time for each learning step were compared. In addition, visual scoring was considered to comparatively evaluate the manual and automatic segmentation. Results After learning, the DSC gradually increased to 0.48 ± 0.11 to 0.50 ± 0.11, and 0.58 ± 0.08. The DSC for the external dataset was 0.49 ± 0.12. The times required for segmentation were 124.8, 143.4, and 86.4 s, showing a large decrease at the final stage. In visual scoring, the accuracy of manual segmentation was found to be higher than that of automatic segmentation. Conclusions The deep active learning framework can serve as a fast, accurate, and robust clinical tool for demarcating IAN location.

Quantitative and Semi-quantitative Methods for Assessing the Degree of Methylene Blue Staining in Sentinel Lymph Nodes in Dogs

Background: To develop a digital algorithm for quantitative assessment of surface methylene blue staining in whole lymph nodes and validate a semi-quantitative visual scoring method for patient-side use.Methods: Lymph nodes from canine patients with spontaneous tumors undergoing sentinel lymph node mapping were prospectively assessed ex vivo and photographed. Using an open-source computer-based imaging software, an algorithm was developed for quantification of staining based on a signal-to-background ratio. Next, two blinded observers evaluated images and assigned a semi-quantitative visual score based on surface staining (0—no blue stain, 1−1–50% stained, and 2−51–100% stained) and those results were compared to the established quantitative standard.Results: Forty-three lymph nodes were included. Image analysis successfully quantified blue staining and differentiated from normal lymph node tissue in all cases. Agreement between observers using the Kappa coefficient demonstrated strong agreement (k = 0.8581, p < 0.0001) between semi-quantitative visual scoring and image analysis. There was substantial interobserver and intraobserver agreement for the scoring system (k = 0.7340, p < 0.0001 and k = 0.8983, p < 0.0001, respectively).Conclusion: A digital algorithm using an open-source software was simple and straightforward to use for quantification of blue staining. The use of a semi-quantitative visual scoring system shows promise for a simple, objective, repeatable assessment of methylene blue staining at the time of surgery. This study demonstrates reliable and repeatable methods for blue staining quantification thereby providing a novel and objective reporting mechanism in scientific research involving sentinel lymph node mapping.

An open-source, high-performance tool for automated sleep staging

The clinical and societal measurement of human sleep has increased exponentially in recent years. However, unlike other fields of medical analysis that have become highly automated, basic and clinical sleep research still relies on human visual scoring. Such human-based evaluations are time-consuming, tedious, and can be prone to subjective bias. Here, we describe a novel algorithm trained and validated on +30,000 hr of polysomnographic sleep recordings across heterogeneous populations around the world. This tool offers high sleep-staging accuracy that matches human scoring accuracy and interscorer agreement no matter the population kind. The software is designed to be especially easy to use, computationally low-demanding, open source, and free. Our hope is that this software facilitates the broad adoption of an industry-standard automated sleep staging software package.

Beyond traditional visual sleep scoring: massive feature extraction and unsupervised clustering of sleep time series

Sleep is classically measured with electrophysiological recordings, which are then scored based on guidelines tailored for the visual inspection of these recordings. As such, these rules reflect a limited range of features easily captured by the human eye and do not always reflect the physiological changes associated with sleep. Here we present a novel analysis framework that characterizes sleep using over 7700 time-series features from the hctsa software. We used clustering to categorize sleep epochs based on the similarity of their features, without relying on established scoring conventions. The resulting structure overlapped substantially with that defined by visual scoring and we report novel features that are highly discriminative of sleep stages. However, we also observed discrepancies as hctsa features unraveled distinctive properties within traditional sleep stages. Our framework lays the groundwork for a data-driven exploration of sleep and the identification of new signatures of sleep disorders and conscious sleep states.

A novel CT scoring method predicts the prognosis of interstitial lung disease associated with anti-MDA5 positive dermatomyositis

Anti-melanoma differentiation-associated gene 5-positive dermatomyositis-associated interstitial lung disease (MDA5+ DM-ILD) is a life-threatening disease. This study aimed to develop a novel pulmonary CT visual scoring method for assessing the prognosis of the disease, and an artificial intelligence (AI) algorithm-based analysis and an idiopathic pulmonary fibrosis (IPF)-based scoring were conducted as comparators. A retrospective cohort of hospitalized patients with MDA5+ DM-ILD was analyzed. Since most fatalities occur within the first half year of the disease course, the primary outcome was the six-month all-cause mortality since the time of admission. A ground glass opacity (GGO) and consolidation-weighted CT visual scoring model for MDA5+ DM-ILD, namely ‘MDA5 score’, was then developed with C-index values of 0.80 (95%CI 0.75–0.86) in the derivation dataset (n = 116) and 0.84 (95%CI 0.71–0.97) in the validation dataset (n = 57), respectively. While, the AI algorithm-based analysis, namely ‘AI score’, yielded C-index 0.78 (95%CI 0.72–0.84) for the derivation dataset and 0.77 (95%CI 0.64–0.90) for the validation dataset. These findings suggest that the newly derived ‘MDA5 score’ may serve as an applicable prognostic predictor for MDA5+ DM-ILD and facilitate further clinical trial design. The AI based CT quantitative analysis provided a promising solution for ILD evaluation.

An optimized imaging protocol for [99mTc]Tc-DPD scintigraphy and SPECT/CT quantification in cardiac transthyretin (ATTR) amyloidosis

Background In [99mTc]Tc-DPD scintigraphy for myocardial ATTR amyloidosis, planar images 3 hour p.i. and SPECT/CT acquisition in L-mode are recommended. This study investigated if earlier planar images (1 hour p.i.) are beneficial and if SPECT/CT acquisition should be preferred in H-mode (180° detector angle) or L-mode (90°). Methods In SPECT/CT phantom measurements (NaI cameras, N = 2; CZT, N = 1), peak contrast recovery (CRpeak) was derived from sphere inserts or myocardial insert (cardiac phantom; signal-to-background ratio [SBR], 10:1 or 5:1). In 25 positive and 38 negative patients (reference: endomyocardial biopsy or clinical diagnosis), Perugini scores and heart-to-contralateral (H/CL) count ratios were derived from planar images 1 hour and 3 hour p.i. Results In phantom measurements, accuracy of myocardial CRpeak at SBR 10:1 (H-mode, 0.95-0.99) and reproducibility at 5:1 (H-mode, 1.02-1.14) was comparable for H-mode and L-mode. However, L-mode showed higher variability of background counts and sphere CRpeak throughout the field of view than H-mode. In patients, sensitivity/specificity were ≥ 95% for H/CL ratios at both time points and visual scoring 3 hour. At 1 hour, visual scores showed specificity of 89% and reduced reader’s confidence. Conclusions Early DPD images provided no additional value for visual scoring or H/CL ratios. In SPECT/CT, H-mode is preferred over L-mode, especially if quantification is applied apart from the myocardium.

A single-center comparative study of lung ultrasound versus chest computed tomography during the COVID-19 era

Background: Lung ultrasound (LUS) is a bedside imaging tool that has proven useful in identifying and assessing the severity of pulmonary pathology. The aim of this study was to determine LUS patterns, their clinical significance, and how they compare to CT findings in hospitalized patients with coronavirus infection.Methods: This observational study included 62 patients (33 men, age 59.3±15.9 years), hospitalized with pneumonia due to COVID-19, who underwent chest CT and bedside LUS on the day of admission. The CT images were analyzed by chest radiographers who calculated a CT visual score based on the expansion and distribution of ground-glass opacities and consolidations. The LUS score was calculated according to the presence, distribution, and severity of anomalies.Results: All patients had CT findings suggestive of bilateral COVID-19 pneumonia, with an average visual scoring of 8.1±2.9%. LUS identified 4 different abnormalities, with bilateral distribution (mean LUS score: 26.4±6.7), focal areas of non-confluent B lines, diffuse confluent B lines, small sub-pleural micro consolidations with pleural line irregularities, and large parenchymal consolidations with air bronchograms. LUS score was significantly correlated with CT visual scoring (rho = 0.70; p<0.001). Correlation analysis of the CT and LUS severity scores showed good interclass correlation (ICC) (ICC =0.71; 95% confidence interval (CI): 0.52–0.83; p<0.001). Logistic regression was used to determine the cut-off value of ≥27 (area under the curve: 0.97; 95% CI: 90-99; sensitivity 88.5% and specificity 97%) of the LUS severity score that represented severe and critical pulmonary involvement on chest CT (CT: 3-4).Conclusion: When combined with clinical data, LUS can provide a potent diagnostic aid in patients with suspected COVID-19 pneumonia, reflecting CT findings.

Deep Semi-Supervised Learning for Automatic Segmentation of Inferior Alveolar Nerve Using a Convolutional Neural Network (Preprint)

BACKGROUND The inferior alveolar nerve (IAN) innervates and regulates the sensation of the mandibular teeth and lower lip. The position of the IAN should be monitored during surgery to prevent damage. Therefore, a study using artificial intelligence (AI) was planned to image and track the position of the IAN automatically for a quicker and safer surgery. OBJECTIVE In this study, we segmented the precise position of the IAN using AI. The accuracy of this technique was evaluated by comparing the position with the position manually specified by a specialist, and segmentation accuracy and annotation efficiency were found to be improved with learning. METHODS A total of 138 cone-beam computed tomography datasets (Internal: 98, External: 40) collected from multiple centers (three hospitals) were used in the study. A customized 3D nnU-Net was used for image segmentation. Active learning, which consists of three steps, was carried out in iterations for 83 datasets with cumulative additions after each step. Subsequently, the accuracy of the model for IAN segmentation was evaluated using the residual dataset. We compared the accuracy by deriving the dice similarity coefficient (DSC) value and the segmentation time for each learning step. In addition, visual scoring was considered to comparatively evaluate the manual and automatic segmentation. RESULTS After learning, the DSC gradually increased to 0.48 ± 0.11 to 0.50 ± 0.11, and 0.58 ± 0.08. The DSC for the external dataset was 0.49 ± 0.12. The times required for segmentation were 124.8, 143.4, and 86.4 s, showing a large decrease at the final stage. In visual scoring, the accuracy of manual segmentation was found to be higher than that of automatic segmentation. CONCLUSIONS The deep active learning framework can serve as a fast, accurate, and robust clinical tool for demarcating IAN location.

Frequency of flow limitation using airflow shape

Study Objectives The presence of flow limitation during sleep is associated with adverse health consequences independent of obstructive sleep apnea (OSA) severity (apnea-hypopnea index, AHI), but remains extremely challenging to quantify. Here we present a unique library and an accompanying automated method that we apply to investigate flow limitation during sleep. Methods A library of 117,871 breaths (N=40 participants) were visually classified (certain flow limitation, possible flow limitation, normal) using airflow shape and physiological signals (ventilatory drive per intra-esophageal diaphragm EMG). An ordinal regression model was developed to quantify flow limitation certainty using flow-shape features (e.g. flattening, scooping); breath-by-breath agreement (Cohen’s ƙ) and overnight flow limitation frequency (R 2, %breaths in certain or possible categories during sleep) were compared against visual scoring. Subsequent application examined flow limitation frequency during arousals and stable breathing, and associations with ventilatory drive. Results The model (23 features) assessed flow limitation with good agreement (breath-by-breath ƙ=0.572, p&lt;0.001) and minimal error (overnight flow limitation frequency R 2=0.86, error=7.2%). Flow limitation frequency was largely independent of AHI (R 2=0.16) and varied widely within individuals with OSA (74[32-95]%breaths, mean[range], AHI&gt;15/hr, N=22). Flow limitation was unexpectedly frequent but variable during arousals (40[5-85]%breaths) and stable breathing (58[12-91]%breaths), and was associated with elevated ventilatory drive (R 2=0.26-0.29; R 2&lt;0.01 AHI v. drive). Conclusions Our method enables quantification of flow limitation frequency, a key aspect of obstructive sleep-disordered breathing that is independent of the AHI and often unavailable. Flow limitation frequency varies widely between individuals, is prevalent during arousals and stable breathing, and reveals elevated ventilatory drive.