Deep Network-Based Comprehensive Parotid Gland Tumor Detection

The heart sound classification of congenital heart disease by using median EEMD-Hurst and threshold denoising method

Heart sound signals are vital for the machine-assisted detection of congenital heart disease. However, the performance of diagnostic results is limited by noise during heart sound acquisition. A limitation of existing noise reduction schemes is that the pathological components of the signal are weak, which have the potential to be filtered out with the noise. In this research, a novel approach for classifying heart sounds based on median ensemble empirical mode decomposition (MEEMD), Hurst analysis, improved threshold denoising, and neural networks are presented. In decomposing the heart sound signal into several intrinsic mode functions (IMFs), mode mixing and mode splitting can be effectively suppressed by MEEMD. Hurst analysis is adopted for identifying the noisy content of IMFs. Then, the noise-dominated IMFs are denoised by an improved threshold function. Finally, the noise reduction signal is generated by reconstructing the processed components and the other components. A database of 5000 heart sounds from congenital heart disease and normal volunteers was constructed. The Mel spectral coefficients of the denoised signals were used as input vectors to the convolutional neural network for classification to verify the effectiveness of the preprocessing algorithm. An accuracy of 93.8%, a specificity of 93.1%, and a sensitivity of 94.6% were achieved for classifying the normal cases from abnormal one.

Tissue specific tumor-gene link prediction through sampling based GNN using a heterogeneous network

A tissue sample is a valuable resource for understanding a patient's symptoms and health status in relation to tumor growth. Recent research seeks to establish a connection between tissue-specific tumor samples and genetic markers (genes). This breakthrough has paved the way for personalized cancer therapies. With this motivation, the proposed model constructs a heterogeneous network based on tumor sample-gene relation data and gene-gene interaction data. This network also incorporates tissue-specific gene expression and primary site-based gene counts as features, enabling tissue-specific predictions. Graph neural networks (GNNs) have proven effective in modeling complex interactions and predicting links within this network. The proposed model has successfully predicted tumor-gene associations by leveraging sampling-based GNNs and link layer embedding. The model's performance metrics, such as AUC-ROC scores, reached approximately 94%, demonstrating the potential of this heterogeneous network in predicting tissue-specific tumor sample-gene links. This paper's findings highlight the importance of tissue-specific associations in cancer research.

Segmentation-Free Outcome Prediction from Head and Neck Cancer PET/CT Images: Deep Learning-Based Feature Extraction from Multi-Angle Maximum Intensity Projections (MA-MIPs)

We introduce an innovative, simple, effective segmentation-free approach for survival analysis of head and neck cancer (HNC) patients from PET/CT images. By harnessing deep learning-based feature extraction techniques and multi-angle maximum intensity projections (MA-MIPs) applied to Fluorodeoxyglucose Positron Emission Tomography (FDG-PET) images, our proposed method eliminates the need for manual segmentations of regions-of-interest (ROIs) such as primary tumors and involved lymph nodes. Instead, a state-of-the-art object detection model is trained utilizing the CT images to perform automatic cropping of the head and neck anatomical area, instead of only the lesions or involved lymph nodes on the PET volumes. A pre-trained deep convolutional neural network backbone is then utilized to extract deep features from MA-MIPs obtained from 72 multi-angel axial rotations of the cropped PET volumes. These deep features extracted from multiple projection views of the PET volumes are then aggregated and fused, and employed to perform recurrence-free survival analysis on a cohort of 489 HNC patients. The proposed approach outperforms the best performing method on the target dataset for the task of recurrence-free survival analysis. By circumventing the manual delineation of the malignancies on the FDG PET-CT images, our approach eliminates the dependency on subjective interpretations and highly enhances the reproducibility of the proposed survival analysis method. The code for this work is publicly released.

SymTC: A symbiotic Transformer-CNN net for instance segmentation of lumbar spine MRI

Intervertebral disc disease, a prevalent ailment, frequently leads to intermittent or persistent low back pain, and diagnosing and assessing of this disease rely on accurate measurement of vertebral bone and intervertebral disc geometries from lumbar MR images. Deep neural network (DNN) models may assist clinicians with more efficient image segmentation of individual instances (discs and vertebrae) of the lumbar spine in an automated way, which is termed as instance image segmentation. In this work, we proposed SymTC, an innovative lumbar spine MR image segmentation model that combines the strengths of Transformer and Convolutional Neural Network (CNN). Specifically, we designed a parallel dual-path architecture to merge CNN layers and Transformer layers, and we integrated a novel position embedding into the self-attention module of Transformer, enhancing the utilization of positional information for more accurate segmentation. To further improve model performance, we introduced a new data synthesis technique to create synthetic yet realistic MR image dataset, named SSMSpine, which is made publicly available. We evaluated our SymTC and the other 16 representative image segmentation models on our private in-house dataset and public SSMSpine dataset, using two metrics, Dice Similarity Coefficient and the 95th percentile Hausdorff Distance. The results indicate that SymTC surpasses the other 16 methods, achieving the highest dice score of 96.169 % for segmenting vertebral bones and intervertebral discs on the SSMSpine dataset. The SymTC code and SSMSpine dataset are publicly available at https://github.com/jiasongchen/SymTC.

Multi-file dynamic compression method based on classification algorithm in DNA storage

The exponential growth in data volume has necessitated the adoption of alternative storage solutions, and DNA storage stands out as the most promising solution. However, the exorbitant costs associated with synthesis and sequencing impeded its development. Pre-compressing the data is recognized as one of the most effective approaches for reducing storage costs. However, different compression methods yield varying compression ratios for the same file, and compressing a large number of files with a single method may not achieve the maximum compression ratio. This study proposes a multi-file dynamic compression method based on machine learning classification algorithms that selects the appropriate compression method for each file to minimize the amount of data stored into DNA as much as possible. Firstly, four different compression methods are applied to the collected files. Subsequently, the optimal compression method is selected as a label, as well as the file type and size are used as features, which are put into seven machine learning classification algorithms for training. The results demonstrate that k-nearest neighbor outperforms other machine learning algorithms on the validation set and test set most of the time, achieving an accuracy rate of over 85% and showing less volatility. Additionally, the compression rate of 30.85% can be achieved according to k-nearest neighbor model, more than 4.5% compared to the traditional single compression method, resulting in significant cost savings for DNA storage in the range of $0.48 to 3 billion/TB. In comparison to the traditional compression method, the multi-file dynamic compression method demonstrates a more significant compression effect when compressing multiple files. Therefore, it can considerably decrease the cost of DNA storage and facilitate the widespread implementation of DNA storage technology.

Predicting High-risk Capsular Features in Pleomorphic Adenoma of the Parotid Gland Through a Nomogram Model Based on ADC Imaging

RATIONALE AND OBJECTIVES

Based on Apparent Diffusion Coefficient (ADC) images, a nomogram model is established to accurately predict the high-risk capsular characteristics associated with pleomorphic adenoma of the parotid gland (PAP) recurrence.

MATERIALS AND METHODS

This retrospective study analyzed 190 patients with PAPs. Significant clinical radiological factors were identified through univariate difference analysis and multivariate regression analysis. The optimal threshold was determined by analyzing the average ADC value of the entire tumor, using the best Youden index and sensitivity analysis, and tumor subregions were delineated accordingly. Three radiomic models were constructed for the whole tumor and for high/low ADC areas, with the best model determined through statistical analysis. Ultimately, a nomogram model was constructed by combining the independent predictive factor of high-risk capsular features with the optimal radiomic predictive score. Model performance was comprehensively assessed by the area under the receiver operating characteristic curve (ROC AUC), accuracy, sensitivity, and specificity.

RESULTS

The best ADC division threshold as 1.25 × 10-3 mm2/s. Multivariate analysis identified High-ADC Zone Volume Percentage as an independent predictor for PAPs with high-risk capsular characteristics. The radiomic model based on the low ADC tumor subregion was optimal (AUC 0.899). The nomogram model, combining independent predictors and optimal imaging studies predictive score, demonstrated high performance (AUC 0.909). Decision curve analysis confirmed the nomogram's clinical applicability.

CONCLUSION

The nomogram model constructed from ADC quantitative imaging can predict PAPs patients with high-risk capsular features. These patients require intraoperative preventive measures to avoid tumor spillage and residuals, as well as extended postoperative follow-up.

A unified approach for Parkinson's disease recognition: imbalance mitigation and grid search optimized boosting with LightGBM

The work elucidates the importance of accurate Parkinson's disease classification within medical diagnostics and introduces a novel framework for achieving this goal. Specifically, the study focuses on enhancing disease identification accuracy utilizing boosting methods. A standout contribution of this work lies in the utilization of a light gradient boosting machine (LGBM) coupled with hyperparameter tuning through grid search optimization (GSO) on the Parkinson's disease dataset derived from speech recording signals. In addition, the Synthetic Minority Over-sampling Technique (SMOTE) has also been employed as a pre-processing technique to balance the dataset, enhancing the robustness and reliability of the analysis. This approach is a novel addition to the study and underscores its potential to enhance disease identification accuracy. The datasets employed in this work include both gender-specific and combined cases, utilizing several distinctive feature subsets including baseline, Mel-frequency cepstral coefficients (MFCC), time-frequency, wavelet transform (WT), vocal fold, and tunable-Q-factor wavelet transform (TQWT). Comparative analyses against state-of-the-art boosting methods, such as AdaBoost and XG-Boost, reveal the superior performance of our proposed approach across diverse datasets and metrics. Notably, on the male cohort dataset, our method achieves exceptional results, demonstrating an accuracy of 0.98, precision of 1.00, sensitivity of 0.97, F1-Score of 0.98, and specificity of 1.00 when utilizing all features with GSO-LGBM. In comparison to AdaBoost and XGBoost, the proposed framework utilizing LGBM demonstrates superior accuracy, achieving an average improvement of 5% in classification accuracy across all feature subsets and datasets. These findings underscore the potential of the proposed methodology to enhance disease identification accuracy and provide valuable insights for further advancements in medical diagnostics.

Cancer detection and classification using a simplified binary state vector machine

Cancer is an invasive and malignant growth of cells and is known to be one of the most fatal diseases. Its early detection is essential for decreasing the mortality rate and increasing the probability of survival. This study presents an efficient machine learning approach based on the state vector machine (SVM) to diagnose and classify tumors into malignant or benign cancer using the online lymphographic data. Further, two types of neural network architectures are also implemented to evaluate the performance of the proposed SVM-based approach. The optimal structures of the classifiers are obtained by varying the architecture, topology, learning rate, and kernel function and recording the results' accuracy. The classifiers are trained with the preprocessed data examples after noise removal and tested on the unknown cases to diagnose each example as positive or negative. Further, the positive cases are classified into different stages including metastases, malign lymph, and fibrosis. The results are evaluated against the feed-forward and generalized regression neural networks. It is found that the proposed SVM-based approach significantly improves the early detection and classification accuracy in comparison to the experienced physicians and the other machine learning approaches. The proposed approach is robust and can perform sub-class divisions for multipurpose tasks. Experimental results demonstrate that the two-class SVM gives the best results and can effectively be used for the classification of cancer. It has outperformed all other classifiers with an average accuracy of 94.90%.

Evaluation of Quantitative Dual-Energy Computed Tomography Parameters for Differentiation of Parotid Gland Tumors

RATIONALE AND OBJECTIVES

To assess the diagnostic performance of quantitative parameters from dual-energy CT (DECT) in differentiating parotid gland tumors (PGTs).

MATERIALS AND METHODS

101 patients with 108 pathologically proved PGTs were enrolled and classified into four groups: pleomorphic adenomas (PAs), warthin tumors (WTs), other benign tumors (OBTs), and malignant tumors (MTs). Conventional CT attenuation and DECT quantitative parameters, including iodine concentration (IC), normalized iodine concentration (NIC), effective atomic number (Zeff), electron density (Rho), double energy index (DEI), and the slope of the spectral Hounsfield unit curve (λHU), were obtained and compared between benign tumors (BTs) and MTs, and further compared among the four subgroups. Logistic regression analysis was used to assess the independent parameters and the receiver operating characteristic (ROC) curves were used to analyze the diagnostic performance.

RESULTS

Attenuation, Zeff, DEI, IC, NIC, and λHU in the arterial phase (AP) and venous phase (VP) were higher in MTs than in BTs (p < 0.001-0.047). λHU in VP and Zeff in AP were independent predictors with an area under the curve (AUC) of 0.84 after the combination. Furthermore, attenuation, Zeff, DEI, IC, NIC, and λHU in the AP and VP of MTs were higher than those of PAs (p < 0.001-0.047). Zeff and NIC in AP and λHU in VP were independent predictors with an AUC of 0.93 after the combination. Attenuation and Rho in the precontrast phase; attenuation, Rho, Zeff, DEI, IC, NIC, and λHU in AP; and the Rho in the VP of PAs were lower than those of WTs (p < 0.001-0.03). Rho in the precontrast phase and attenuation in AP were independent predictors with an AUC of 0.89 after the combination. MTs demonstrated higher Zeff, DEI, IC, NIC, and λHU in VP and lower Rho in the precontrast phase compared with WTs (p < 0.001-0.04); but no independent predictors were found.

CONCLUSION

DECT quantitative parameters can help to differentiate PGTs.

Masked autoencoders with generalizable self-distillation for skin lesion segmentation

In the field of skin lesion image segmentation, accurate identification and partitioning of diseased regions is of vital importance for in-depth analysis of skin cancer. Self-supervised learning, i.e., MAE, has emerged as a potent force in the medical imaging domain, which autonomously learns and extracts latent features from unlabeled data, thereby yielding pre-trained models that greatly assist downstream tasks. To encourage pre-trained models to more comprehensively learn the global structural and local detail information inherent in dermoscopy images, we introduce a Teacher-Student architecture, named TEDMAE, by incorporating a self-distillation mechanism, it learns holistic image feature information to improve the generalizable global knowledge learning of the student MAE model. To make the image features learned by the model suitable for unknown test images, two optimization strategies are, Exterior Conversion Augmentation (EC) utilizes random convolutional kernels and linear interpolation to effectively transform the input image into one with the same shape but altered intensities and textures, while Dynamic Feature Generation (DF) employs a nonlinear attention mechanism for feature merging, enhancing the expressive power of the features, are proposed to enhance the generalizability of global features learned by the teacher model, thereby improving the overall generalization capability of the pre-trained models. Experimental results from the three public skin disease datasets, ISIC2019, ISIC2017, and PH2 indicate that our proposed TEDMAE method outperforms several similar approaches. Specifically, TEDMAE demonstrated optimal segmentation and generalization performance on the ISIC2017 and PH2 datasets, with Dice scores reaching 82.1% and 91.2%, respectively. The best Jaccard values were 72.6% and 84.5%, while the optimal HD95% values were 13.0% and 8.9%, respectively.

SALW-Net: a lightweight convolutional neural network based on self-adjusting loss function for spine MR image segmentation

Segmentation of intervertebral discs and vertebrae from spine magnetic resonance (MR) images is essential to aid diagnosis algorithms for lumbar disc herniation. Convolutional neural networks (CNN) are effective methods, but often require high computational costs. Designing a lightweight CNN is more suitable for medical sites lacking high-computing power devices, yet due to the unbalanced pixel distribution in spine MR images, the segmentation is often sub-optimal. To address this issue, a lightweight spine segmentation CNN based on a self-adjusting loss function, which is named SALW-Net, is proposed in this study. For SALW-Net, the self-adjusting loss function could dynamically adjust the loss weights of the two branches according to the differences in segmentation results and labels during the training; thus, the ability for learning unbalanced pixels is enhanced. Two separate datasets are used to evaluate the proposed SALW-Net. Specifically, the proposed SALW-Net has fewer parameter numbers than U-net (only 2%) but achieves higher evaluation scores than that of U-net (the average DSC score of SALW-Net is 0.8781, and that of U-net is 0.8482). In addition, the practicality validation for SALW-Net is also proceeding, including deploying the model on a lightweight device and producing an aid diagnosis algorithm based on segmentation results. This means our SALW-Net has clinical application potential for assisted diagnosis in low computational power scenarios.

MCAFNet: multiscale cross-layer attention fusion network for honeycomb lung lesion segmentation

Accurate segmentation of honeycomb lung lesions from lung CT images plays a crucial role in the diagnosis and treatment of various lung diseases. However, the availability of algorithms for automatic segmentation of honeycomb lung lesions remains limited. In this study, we propose a novel multi-scale cross-layer attention fusion network (MCAFNet) specifically designed for the segmentation of honeycomb lung lesions, taking into account their shape specificity and similarity to surrounding vascular shadows. The MCAFNet incorporates several key modules to enhance the segmentation performance. Firstly, a multiscale aggregation (MIA) module is introduced in the input part to preserve spatial information during downsampling. Secondly, a cross-layer attention fusion (CAF) module is proposed to capture multiscale features by integrating channel information and spatial information from different layers of the feature maps. Lastly, a bidirectional attention gate (BAG) module is constructed within the skip connection to enhance the model's ability to filter out background information and focus on the segmentation target. Experimental results demonstrate the effectiveness of the proposed MCAFNet. On the honeycomb lung segmentation dataset, the network achieves an Intersection over Union (IoU) of 0.895, mean IoU (mIoU) of 0.921, and mean Dice coefficient (mDice) of 0.949, outperforming existing medical image segmentation algorithms. Furthermore, experiments conducted on additional datasets confirm the generalizability and robustness of the proposed model. The contribution of this study lies in the development of the MCAFNet, which addresses the lack of automated segmentation algorithms for honeycomb lung lesions. The proposed network demonstrates superior performance in accurately segmenting honeycomb lung lesions, thereby facilitating the diagnosis and treatment of lung diseases. This work contributes to the existing literature by presenting a novel approach that effectively combines multi-scale features and attention mechanisms for lung lesion segmentation. The code is available at https://github.com/Oran9er/MCAFNet .

When deep learning is not enough: artificial life as a supplementary tool for segmentation of ultrasound images of breast cancer

Segmentation of tumors in ultrasound (US) images of the breast is a critical issue in medical imaging. Due to the poor quality of US images and the varying specifications of US machines, segmentation and classification of abnormalities present difficulties even for trained radiologists. The paper aims to introduce a novel AI-based hybrid model for US segmentation that offers high accuracy, requires relatively smaller datasets, and is capable of handling previously unseen data. The software can be used for diagnostics and the US-guided biopsies. A unique and robust hybrid approach that combines deep learning (DL) and multi-agent artificial life (AL) has been introduced. The algorithms are verified on three US datasets. The method outperforms 14 selected state-of-the-art algorithms applied to US images characterized by complex geometry and high level of noise. The paper offers an original classification of the images and tests to analyze the limits of the DL. The model has been trained and verified on 1264 ultrasound images. The images are in the JPEG and PNG formats. The age of the patients ranges from 22 to 73 years. The 14 benchmark algorithms include deformable shapes, edge linking, superpixels, machine learning, and DL methods. The tests use eight-region shape- and contour-based evaluation metrics. The proposed method (DL-AL) produces excellent results in terms of the dice coefficient (region) and the relative Hausdorff distance H3 (contour-based) as follows: the easiest image complexity level, Dice = 0.96 and H3 = 0.26; the medium complexity level, Dice = 0.91 and H3 = 0.82; and the hardest complexity level, Dice = 0.90 and H3 = 0.84. All other metrics follow the same pattern. The DL-AL outperforms the second best (Unet-based) method by 10-20%. The method has been also tested by a series of unconventional tests. The model was trained on low complexity images and applied to the entire set of images. These results are summarized below. (1) Only the low complexity images have been used for training (68% unknown images): Dice = 0.80 and H3 = 2.01. (2) The low and the medium complexity images have been used for training (51% unknown images): Dice = 0.86 and H3 = 1.32. (3) The low, medium, and hard complexity images have been used for training (35% unknown images): Dice = 0.92 and H3 = 0.76. These tests show a significant advantage of DL-AL over 30%. A video demo illustrating the algorithm is at http://tinyurl.com/mr4ah687 .

EEGNet-based multi-source domain filter for BCI transfer learning

Deep learning has great potential on decoding EEG in brain-computer interface. While common deep learning algorithms cannot directly train models with data from multiple individuals because of the inter-individual differences in EEG. Collecting enough data for each subject to satisfy the training of deep learning would result in an increase in training cost. This study proposes a novel transfer learning, EEGNet-based multi-source domain filter for transfer learning (EEGNet-MDFTL), to reduce the amount of training data and improve the performance of BCI. The EEGNet-MDFTL uses bagging ensemble learning to learn domain-invariant features from the multi-source domain and utilizes model loss value to filter the multi-source domain. Compared with baseline methods, the accuracy of the EEGNet-MDFTL reaches 91.96%, higher than two state-of-the-art methods, which demonstrates source domain filter can select similar source domains to improve the accuracy of the model, and remains a high level even when the data amount is reduced to 1/8, proving that ensemble learning learns enough domain invariant features from the multi-source domain to make the model insensitive to data amount. The proposed EEGNet-MDFTL is effective in improving the decoding performance with a small amount of data, which is helpful to save the BCI training cost.

Double-sided asymmetric method for automated fetal heart rate baseline calculation

The fetal heart rate (FHR) signal is used to assess the well-being of a fetus during labor. Manual interpretation of the FHR is subject to high inter- and intra-observer variability, leading to inconsistent clinical decision-making. The baseline of the FHR signal is crucial for its interpretation. An automated method for baseline determination may reduce interpretation variability. Based on this claim, we present the Auto-Regressed Double-Sided Improved Asymmetric Least Squares (ARDSIAsLS) method as a baseline calculation algorithm designed to imitate expert obstetrician baseline determination. As the FHR signal is prone to a high rate of missing data, a step of gap interpolation in a physiological manner was implemented in the algorithm. The baseline of the interpolated signal was determined using a weighted algorithm of two improved asymmetric least squares smoothing models and an improved symmetric least squares smoothing model. The algorithm was validated against a ground truth determined from annotations of six expert obstetricians. FHR baseline calculation performance of the ARDSIAsLS method yielded a mean absolute error of 2.54 bpm, a max absolute error of 5.22 bpm, and a root mean square error of 2.89 bpm. In a comparison between the algorithm and 11 previously published methods, the algorithm outperformed them all. Notably, the algorithm was non-inferior to expert annotations. Automating the baseline FHR determination process may help reduce practitioner discordance and aid decision-making in the delivery room.

A novel computer modeling and simulation technique for bronchi motion tracking in human lungs under respiration

In this work, we proposed a novel computer modeling and simulation technique for motion tracking of lung bronchi (or tumors) under respiration using 9 cases of computed tomography (CT)-based patient-specific finite element (FE) models and Ogden's hyperelastic model. In the fabrication of patient-specific FE models for the respiratory system, various organs such as the mediastinum, diaphragm, and thorax that could affect the lung motions during breathing were considered. To describe the nonlinear material behavior of lung parenchyma, the comparative simulation for biaxial tension-compression of lung parenchyma was carried out using several hyperelastic models in ABAQUS, and then, Ogden's model was adopted as an optimal model. Based on the aforementioned FE models and Ogden's material model, the 9 cases of respiration simulation were carried out from exhalation to inhalation, and the motion of lung bronchi (or tumors) was tracked. In addition, the changes in lung volume, lung cross-sectional area on the axial plane during breathing were calculated. Finally, the simulation results were quantitatively compared to the inhalation/exhalation CT images of 9 subjects to validate the proposed technique. Through the simulation, it was confirmed that the average relative errors of simulation to clinical data regarding to the displacement of 258 landmarks in the lung bronchi branches of total subjects were 1.10%~2.67%. In addition, the average relative errors of those with respect to the lung cross-sectional area changes and the volume changes in the superior-inferior direction were 0.20%~5.00% and 1.29 ~ 9.23%, respectively. Hence, it was considered that the simulation results were coincided well with the clinical data. The novelty of the present study is as follows: (1) The framework from fabrication of the human respiratory system to validation of the bronchi motion tracking is provided step by step. (2) The comparative simulation study for nonlinear material behavior of lung parenchyma was carried out to describe the realistic lung motion. (3) Various organs surrounding the lung parenchyma and restricting its motion were considered in respiration simulation. (4) The simulation results such as landmark displacement, lung cross-sectional area/volume changes were quantitatively compared to the clinical data of 9 subjects.

ECG-based convolutional neural network in pediatric obstructive sleep apnea diagnosis

Obstructive sleep apnea (OSA) is a prevalent respiratory condition in children and is characterized by partial or complete obstruction of the upper airway during sleep. The respiratory events in OSA induce transient alterations of the cardiovascular system that ultimately can lead to increased cardiovascular risk in affected children. Therefore, a timely and accurate diagnosis is of utmost importance. However, polysomnography (PSG), the standard diagnostic test for pediatric OSA, is complex, uncomfortable, costly, and relatively inaccessible, particularly in low-resource environments, thereby resulting in substantial underdiagnosis. Here, we propose a novel deep-learning approach to simplify the diagnosis of pediatric OSA using raw electrocardiogram tracing (ECG). Specifically, a new convolutional neural network (CNN)-based regression model was implemented to automatically predict pediatric OSA by estimating its severity based on the apnea-hypopnea index (AHI) and deriving 4 OSA severity categories. For this purpose, overnight ECGs from 1,610 PSG recordings obtained from the Childhood Adenotonsillectomy Trial (CHAT) database were used. The database was randomly divided into approximately 60%, 20%, and 20% for training, validation, and testing, respectively. The diagnostic performance of the proposed CNN model largely outperformed the most accurate previous algorithms that relied on ECG-derived features (4-class Cohen's kappa coefficient of 0.373 versus 0.166). Specifically, for AHI cutoff values of 1, 5, and 10 events/hour, the binary classification achieved sensitivities of 84.19%, 76.67%, and 53.66%; specificities of 46.15%, 91.39%, and 98.06%; and accuracies of 75.92%, 86.96%, and 91.97%, respectively. Therefore, pediatric OSA can be readily identified by our proposed CNN model, which provides a simpler, faster, and more accessible diagnostic test that can be implemented in clinical practice.

Deep learning system assisted detection and localization of lumbar spondylolisthesis

Objective: Explore a new deep learning (DL) object detection algorithm for clinical auxiliary diagnosis of lumbar spondylolisthesis and compare it with doctors' evaluation to verify the effectiveness and feasibility of the DL algorithm in the diagnosis of lumbar spondylolisthesis. Methods: Lumbar lateral radiographs of 1,596 patients with lumbar spondylolisthesis from three medical institutions were collected, and senior orthopedic surgeons and radiologists jointly diagnosed and marked them to establish a database. These radiographs were randomly divided into a training set (n = 1,117), a validation set (n = 240), and a test set (n = 239) in a ratio of 0.7 : 0.15: 0.15. We trained two DL models for automatic detection of spondylolisthesis and evaluated their diagnostic performance by PR curves, areas under the curve, precision, recall, F1-score. Then we chose the model with better performance and compared its results with professionals' evaluation. Results: A total of 1,780 annotations were marked for training (1,242), validation (263), and test (275). The Faster Region-based Convolutional Neural Network (R-CNN) showed better precision (0.935), recall (0.935), and F1-score (0.935) in the detection of spondylolisthesis, which outperformed the doctor group with precision (0.927), recall (0.892), f1-score (0.910). In addition, with the assistance of the DL model, the precision of the doctor group increased by 4.8%, the recall by 8.2%, the F1-score by 6.4%, and the average diagnosis time per plain X-ray was shortened by 7.139 s. Conclusion: The DL detection algorithm is an effective method for clinical diagnosis of lumbar spondylolisthesis. It can be used as an assistant expert to improve the accuracy of lumbar spondylolisthesis diagnosis and reduce the clinical workloads.