A machine learning-based hybrid recommender framework for smart medical systems

This article presents a hybrid recommender framework for smart medical systems by introducing two methods to improve service level evaluations and doctor recommendations for patients. The first method uses big data techniques and deep learning algorithms to develop a registration review system in medical institutions. This system outperforms conventional evaluation methods, thus achieving higher accuracy. The second method implements the term frequency and inverse document frequency (TF-IDF) algorithm to construct a model based on the patient's symptom vector space, incorporating score weighting, modified cosine similarity, and K-means clustering. Then, the alternating least squares (ALS) matrix decomposition and user collaborative filtering algorithm are applied to calculate patients' predicted scores for doctors and recommend top-performing doctors. Experimental results show significant improvements in metrics called precision and recall rates compared to conventional methods, making the proposed approach a practical solution for department triage and doctor recommendation in medical appointment platforms.

A new framework for water quality forecasting coupling causal inference, time-frequency analysis and uncertainty quantification

Accurate forecasting of water quality variables in river systems is crucial for relevant administrators to identify potential water quality degradation issues and take countermeasures promptly. However, pure data-driven forecasting models are often insufficient to deal with the highly varying periodicity of water quality in today's more complex environment. This study presents a new holistic framework for time-series forecasting of water quality parameters by combining advanced deep learning algorithms (i.e., Long Short-Term Memory (LSTM) and Informer) with causal inference, time-frequency analysis, and uncertainty quantification. The framework was demonstrated for total nitrogen (TN) forecasting in the largest artificial lakes in Asia (i.e., the Danjiangkou Reservoir, China) with six-year monitoring data from January 2017 to June 2022. The results showed that the pre-processing techniques based on causal inference and wavelet decomposition can significantly improve the performance of deep learning algorithms. Compared to the individual LSTM and Informer models, wavelet-coupled approaches diminished well the apparent forecasting errors of TN concentrations, with 24.39%, 32.68%, and 41.26% reduction at most in the average, standard deviation, and maximum values of the errors, respectively. In addition, a post-processing algorithm based on the Copula function and Bayesian theory was designed to quantify the uncertainty of predictions. With the help of this algorithm, each deterministic prediction of our model can correspond to a range of possible outputs. The 95% forecast confidence interval covered almost all the observations, which proves a measure of the reliability and robustness of the predictions. This study provides rich scientific references for applying advanced data-driven methods in time-series forecasting tasks and a practical methodological framework for water resources management and similar projects.

Brain metastasis tumor segmentation and detection using deep learning algorithms: A systematic review and meta-analysis

BACKGROUND

Manual detection of brain metastases is both laborious and inconsistent, driving the need for more efficient solutions. Accordingly, our systematic review and meta-analysis assessed the efficacy of deep learning algorithms in detecting and segmenting brain metastases from various primary origins in MRI images.

METHODS

We conducted a comprehensive search of PubMed, Embase, and Web of Science up to May 24, 2023, which yielded 42 relevant studies for our analysis. We assessed the quality of these studies using the QUADAS-2 and CLAIM tools. Using a random-effect model, we calculated the pooled lesion-wise dice score as well as patient-wise and lesion-wise sensitivity. We performed subgroup analyses to investigate the influence of factors such as publication year, study design, training center of the model, validation methods, slice thickness, model input dimensions, MRI sequences fed to the model, and the specific deep learning algorithms employed. Additionally, meta-regression analyses were carried out considering the number of patients in the studies, count of MRI manufacturers, count of MRI models, training sample size, and lesion number.

RESULTS

Our analysis highlighted that deep learning models, particularly the U-Net and its variants, demonstrated superior segmentation accuracy. Enhanced detection sensitivity was observed with an increased diversity in MRI hardware, both in terms of manufacturer and model variety. Furthermore, slice thickness was identified as a significant factor influencing lesion-wise detection sensitivity. Overall, the pooled results indicated a lesion-wise dice score of 79%, with patient-wise and lesion-wise sensitivities at 86% and 87%, respectively.

CONCLUSIONS

The study underscores the potential of deep learning in improving brain metastasis diagnostics and treatment planning. Still, more extensive cohorts and larger meta-analysis are needed for more practical and generalizable algorithms. Future research should prioritize these areas to advance the field. This study was funded by the Gen. & Mrs. M.C. Peng Fellowship and registered under PROSPERO (CRD42023427776).

DEEP LEARNING ALGORITHMS SHOW SOME POTENTIAL AS AN ADJUNCTIVE TOOL IN CARIES DIAGNOSIS

ARTICLE TITLE AND BIBLIOGRAPHIC INFORMATION

Mohammad-Rahimi H, Reza Motamedian S, Hossein Rohban M, Krois J, Uribe SE, Mahmoudinia E, Rokhshad R, Nadimi M, Schwendicke F, Deep learning for caries detection: A systematic review, J Dent, 2022,122, 104115. ISSN 0300-5712 https://doi.org/10.1016/j.jdent.2022.104115.

SOURCE OF FUNDING

Information not available TYPE OF STUDY/DESIGN: Systematic review.

Street-view greenspace exposure and objective sleep characteristics among children

Greenspace may benefit sleep by enhancing physical activity, reducing stress or air pollution exposure. Studies on greenspace and children's sleep are limited, and most use satellite-derived measures that do not capture ground-level exposures that may be important for sleep. We examined associations of street view imagery (SVI)-based greenspace with sleep in Project Viva, a Massachusetts pre-birth cohort. We used deep learning algorithms to derive novel metrics of greenspace (e.g., %trees, %grass) from SVI within 250m of participant residential addresses during 2007-2010 (mid-childhood, mean age 7.9 years) and 2012-2016 (early adolescence, 13.2y) (N = 533). In early adolescence, participants completed >5 days of wrist actigraphy. Sleep duration, efficiency, and time awake after sleep onset (WASO) were derived from actigraph data. We used linear regression to examine cross-sectional and prospective associations of mid-childhood and early adolescence greenspace exposure with early adolescence sleep, adjusting for confounders. We compared associations with satellite-based greenspace (Normalized Difference Vegetation Index, NDVI). In unadjusted models, mid-childhood SVI-based total greenspace and %trees (per interquartile range) were associated with longer sleep duration at early adolescence (9.4 min/day; 95%CI:3.2,15.7; 8.1; 95%CI:1.7,14.6 respectively). However, in fully adjusted models, only the association between %grass at mid-childhood and WASO was observed (4.1; 95%CI:0.2,7.9). No associations were observed between greenspace and sleep efficiency, nor in cross-sectional early adolescence models. The association between greenspace and sleep differed by racial and socioeconomic subgroups. For example, among Black participants, higher NDVI was associated with better sleep, in neighborhoods with low socio-economic status (SES), higher %grass was associated with worse sleep, and in neighborhoods with high SES, higher total greenspace and %grass were associated with better sleep time. SVI metrics may have the potential to identify specific features of greenspace that affect sleep.

Toward automatic reformation at the orbitomeatal line in head computed tomography using object detection algorithm

Consistent cross-sectional imaging is desirable to accurately detect lesions and facilitate follow-up in head computed tomography (CT). However, manual reformation causes image variations among technologists and requires additional time. We therefore developed a system that reformats head CT images at the orbitomeatal (OM) line and evaluated the system performance using real-world clinical data. Retrospective data were obtained for 681 consecutive patients who underwent non-contrast head CT. The datasets were randomly divided into one of three sets for training, validation, or testing. Four landmarks (bilateral eyes and external auditory canal) were detected with the trained You Look Only Once (YOLO)v5 model, and the head CT images were reformatted at the OM line. The precision, recall, and mean average precision at the intersection over union threshold of 0.5 were computed in the validation sets. The reformation quality in testing sets was evaluated by three radiological technologists on a qualitative 4-point scale. The precision, recall, and mean average precision of the trained YOLOv5 model for all categories were 0.688, 0.949, and 0.827, respectively. In our environment, the mean implementation time was 23.5 ± 2.4 s for each case. The qualitative evaluation in the testing sets showed that post-processed images of automatic reformation had clinically useful quality with scores 3 and 4 in 86.8%, 91.2%, and 94.1% for observers 1, 2, and 3, respectively. Our system demonstrated acceptable quality in reformatting the head CT images at the OM line using an object detection algorithm and was highly time efficient.

A Comparison of Deep Learning Techniques for Arterial Blood Pressure Prediction

Continuous vital signal monitoring is becoming more relevant in preventing diseases that afflict a large part of the world’s population; for this reason, healthcare equipment should be easy to wear and simple to use. Non-intrusive and non-invasive detection methods are a basic requirement for wearable medical devices, especially when these are used in sports applications or by the elderly for self-monitoring. Arterial blood pressure (ABP) is an essential physiological parameter for health monitoring. Most blood pressure measurement devices determine the systolic and diastolic arterial blood pressure through the inflation and the deflation of a cuff. This technique is uncomfortable for the user and may result in anxiety, and consequently affect the blood pressure and its measurement. The purpose of this paper is the continuous measurement of the ABP through a cuffless, non-intrusive approach. The approach of this paper is based on deep learning techniques where several neural networks are used to infer ABP, starting from photoplethysmogram (PPG) and electrocardiogram (ECG) signals. The ABP was predicted first by utilizing only PPG and then by using both PPG and ECG. Convolutional neural networks (ResNet and WaveNet) and recurrent neural networks (LSTM) were compared and analyzed for the regression task. Results show that the use of the ECG has resulted in improved performance for every proposed configuration. The best performing configuration was obtained with a ResNet followed by three LSTM layers: this led to a mean absolute error (MAE) of 4.118 mmHg on and 2.228 mmHg on systolic and diastolic blood pressures, respectively. The results comply with the American National Standards of the Association for the Advancement of Medical Instrumentation. ECG, PPG, and ABP measurements were extracted from the MIMIC database, which contains clinical signal data reflecting real measurements. The results were validated on a custom dataset created at Neuronica Lab, Politecnico di Torino.

A Deep Learning Approach to Re-create Raw Full-Field Digital Mammograms for Breast Density and Texture Analysis

Purpose

To develop a computational approach to re-create rarely stored for-processing (raw) digital mammograms from routinely stored for-presentation (processed) mammograms.

Materials and Methods

In this retrospective study, pairs of raw and processed mammograms collected in 884 women (mean age, 57 years ± 10 [standard deviation]; 3713 mammograms) from October 5, 2017, to August 1, 2018, were examined. Mammograms were split 3088 for training and 625 for testing. A deep learning approach based on a U-Net convolutional network and kernel regression was developed to estimate the raw images. The estimated raw images were compared with the originals by four image error and similarity metrics, breast density calculations, and 29 widely used texture features.

Results

In the testing dataset, the estimated raw images had small normalized mean absolute error (0.022 ± 0.015), scaled mean absolute error (0.134 ± 0.078) and mean absolute percentage error (0.115 ± 0.059), and a high structural similarity index (0.986 ± 0.007) for the breast portion compared with the original raw images. The estimated and original raw images had a strong correlation in breast density percentage (Pearson r = 0.946) and a strong agreement in breast density grade (Cohen κ = 0.875). The estimated images had satisfactory correlations with the originals in 23 texture features (Pearson r ≥ 0.503 or Spearman ρ ≥ 0.705) and were well complemented by processed images for the other six features.

Conclusion

This deep learning approach performed well in re-creating raw mammograms with strong agreement in four image evaluation metrics, breast density, and the majority of 29 widely used texture features.Keywords: Mammography, Breast, Supervised Learning, Convolutional Neural Network (CNN), Deep learning algorithms, Machine Learning AlgorithmsSee also the commentary by Chan in this issue.Supplemental material is available for this article.©RSNA, 2021.

Application of deep autoencoder as an one-class classifier for unsupervised network intrusion detection: a comparative evaluation

The ever-increasing use of internet has opened a new avenue for cybercriminals, alarming the online businesses and organization to stay ahead of evolving thread landscape. To this end, intrusion detection system (IDS) is deemed as a promising defensive mechanism to ensure network security. Recently, deep learning has gained ground in the field of intrusion detection but majority of progress has been witnessed on supervised learning which requires adequate labeled data for training. In real practice, labeling the high volume of network traffic is laborious and error prone. Intuitively, unsupervised deep learning approaches has received gaining momentum. Specifically, the advances in deep learning has endowed autoencoder (AE) with greater ability for data reconstruction to learn the robust feature representation from massive amount of data. Notwithstanding, there is no study that evaluates the potential of different AE variants as one-class classifier for intrusion detection. This study fills this gap of knowledge presenting a comparative evaluation of different AE variants for one-class unsupervised intrusion detection. For this research, the evaluation includes five different variants of AE such as Stacked AE, Sparse AE, Denoising AE, Contractive AE and Convolutional AE. Further, the study intents to conduct a fair comparison establishing a unified network configuration and training scheme for all variants over the common benchmark datasets, NSL-KDD and UNSW-NB15. The comparative evaluation study provides a valuable insight on how different AE variants can be used as one-class classifier to build an effective unsupervised IDS. The outcome of this study will be of great interest to the network security community as it provides a promising path for building effective IDS based on deep learning approaches alleviating the need for adequate and diverse intrusion network traffic behavior.

How much are we exposed to alcohol in electronic media? Development of the Alcoholic Beverage Identification Deep Learning Algorithm (ABIDLA)

BACKGROUND

Evidence demonstrates that seeing alcoholic beverages in electronic media increases alcohol initiation and frequent and excessive drinking, particularly among young people. To efficiently assess this exposure, the aim was to develop the Alcoholic Beverage Identification Deep Learning Algorithm (ABIDLA) to automatically identify beer, wine and champagne/sparkling wine from images.

METHODS

Using a specifically developed software, three coders annotated 57,186 images downloaded from Google. Supplemented by 10,000 images from ImageNet, images were split randomly into training data (70 %), validation data (10 %) and testing data (20 %). For retest reliability, a fourth coder re-annotated a random subset of 2004 images. Algorithms were trained using two state-of-the-art convolutional neural networks, Resnet (with different depths) and Densenet-121.

RESULTS

With a correct classification (accuracy) of 73.75 % when using six beverage categories (beer glass, beer bottle, beer can, wine, champagne, and other images), 84.09 % with three (beer, wine/champagne, others) and 85.22 % with two (beer/wine/champagne, others), Densenet-121 slightly outperformed all Resnet models. The highest accuracy was obtained for wine (78.91 %) followed by beer can (77.43 %) and beer cup (73.56 %). Interrater reliability was almost perfect between the coders and the expert (Kappa = .903) and substantial between Densenet-121 and the coders (Kappa = .681).

CONCLUSIONS

Free from any response or coding burden and with a relatively high accuracy, the ABIDLA offers the possibility to screen all kinds of electronic media for images of alcohol. Providing more comprehensive evidence on exposure to alcoholic beverages is important because exposure instigates alcohol initiation and frequent and excessive drinking.

Anthropogenic Marine Debris assessment with Unmanned Aerial Vehicle imagery and deep learning: A case study along the beaches of the Republic of Maldives

Anthropogenic Marine Debris (AMD) is one of the major environmental issues of our planet to date, and plastic accounts for 80% of total AMD. Beaches represent one of the main marine compartment where AMD accumulates, but few and scattered regional assessments are available from literature reporting quantitative estimation of AMD distributed on the shorelines. However, accessing information on the AMD accumulation rate on beaches, and the associated spatiotemporal oscillations, would be crucial to refining global estimation on the dispersal mechanisms. In our work, we address this issue by proposing an ad-hoc methodology for monitoring and automatically quantifying AMD, based on the combined use of a commercial Unmanned Aerial Vehicle (UAV) (equipped with an RGB high-resolution camera) and a deep-learning based software (i.e.: PlasticFinder). Remote areas were monitored by UAV and were inspected by operators on the ground to check and to categorise all AMD dispersed on the beach. The high-resolution images obtained from UAV allowed to visually detect a percentage of the objects on the shores higher than 87.8%, thus providing suitable images to populate training and testing datasets, as well as gold standards to evaluate the software performance. PlasticFinder reached a Sensitivity of 67%, with a Positive Predictive Value of 94%, in the automatic detection of AMD, but a limitation was found, due to reduced sunlight conditions, thus restricting to the use of the software in its present version. We, therefore, confirmed the efficiency of commercial UAVs as tools for AMD monitoring and demonstrated - for the first time - the potential of deep learning for the automatic detection and quantification of AMD.

IILS: Intelligent imaging layout system for automatic imaging report standardization and intra-interdisciplinary clinical workflow optimization

Background

To achieve imaging report standardization and improve the quality and efficiency of the intra-interdisciplinary clinical workflow, we proposed an intelligent imaging layout system (IILS) for a clinical decision support system-based ubiquitous healthcare service, which is a lung nodule management system using medical images.

Methods

We created a lung IILS based on deep learning for imaging report standardization and workflow optimization for the identification of nodules. Our IILS utilized a deep learning plus adaptive auto layout tool, which trained and tested a neural network with imaging data from all the main CT manufacturers from 11,205 patients. Model performance was evaluated by the receiver operating characteristic curve (ROC) and calculating the corresponding area under the curve (AUC). The clinical application value for our IILS was assessed by a comprehensive comparison of multiple aspects.

Findings

Our IILS is clinically applicable due to the consistency with nodules detected by IILS, with its highest consistency of 0·94 and an AUC of 90·6% for malignant pulmonary nodules versus benign nodules with a sensitivity of 76·5% and specificity of 89·1%. Applying this IILS to a dataset of chest CT images, we demonstrate performance comparable to that of human experts in providing a better layout and aiding in diagnosis in 100% valid images and nodule display. The IILS was superior to the traditional manual system in performance, such as reducing the number of clicks from 14·45 ± 0·38 to 2, time consumed from 16·87 ± 0·38 s to 6·92 ± 0·10 s, number of invalid images from 7·06 ± 0·24 to 0, and missing lung nodules from 46·8% to 0%.

Interpretation

This IILS might achieve imaging report standardization, and improve the clinical workflow therefore opening a new window for clinical application of artificial intelligence.

Fund

The National Natural Science Foundation of China.