Precision lung cancer screening from CT scans using a VGG16-based convolutional neural network.

Freezing of gait (FOG) is one of the most debilitating motor symptoms experienced by patients with Parkinson's disease (PD). FOG detection is possible using acceleration data from wearable sensors, and a convolutional neural network (CNN) is often used to determine the presence of FOG epochs. We compared the performance of a standard CNN for the detection of FOG with two more complex networks, which are well suited for time series data, the MiniRocket and the InceptionTime. We combined acceleration data of people with PD across four studies. The final data set was split into a training (80%) and hold-out test (20%) set. A fifth study was included as an unseen test set. The data were windowed (2 s) and five-fold cross-validation was applied. The CNN, MiniRocket, and InceptionTime models were evaluated using a receiver operating characteristic (ROC) curve and its area under the curve (AUC). Multiple sensor configurations were evaluated for the best model. The geometric mean was subsequently calculated to select the optimal threshold. The selected model and threshold were evaluated on the hold-out and unseen test set. A total of 70 participants (23.7 h, 9% FOG) were included in this study for training and testing, and in addition, 10 participants provided an unseen test set (2.4 h, 11% FOG). The CNN performed best (AUC = 0.86) in comparison to the InceptionTime (AUC = 0.82) and MiniRocket (AUC = 0.76) models. For the CNN, we found a similar performance for a seven-sensor configuration (lumbar, upper and lower legs and feet; AUC = 0.86), six-sensor configuration (upper and lower legs and feet; AUC = 0.87), and two-sensor configuration (lower legs; AUC = 0.86). The optimal threshold of 0.45 resulted in a sensitivity of 77% and a specificity of 58% for the hold-out set (AUC = 0.72), and a sensitivity of 85% and a specificity of 68% for the unseen test set (AUC = 0.90). We confirmed that deep learning can be used to detect FOG in a large, heterogeneous dataset. The CNN model outperformed more complex networks. This model could be employed in future personalized interventions, with the ultimate goal of using automated FOG detection to trigger real-time cues to alleviate FOG in daily life.

An essential noninvasive medical diagnostic technique is magnetic resonance imaging (MRI), which is particularly useful for identifying brain cancers. While earlier algorithms proved effective on smaller MRI datasets, their performance suffered on bigger datasets. This study addresses the need for a swift and reliable brain tumor classification system capable of sustaining optimal performance across comprehensive MRI datasets. The convolutional neural network is implemented using the Keras library, incorporating the ResNet50 architecture as a pre-trained model. The ResNet50 model is fine-tuned for the specific brain tumor classification task, with a Global Average Pooling layer, dropout, and a final dense layer with softmax activation. Data augmentation techniques are employed to enhance the model’s robustness, including rotation, width and height shifts, and horizontal flips. The training process involves optimizing the model using the Adam optimizer with a learning rate of 0.0001. Early stopping, learning rate reduction on plateau, and model checkpointing are implemented as callbacks to ensure efficient training and prevent overfitting. The proposed model achieves a remarkable accuracy of 99.28 percent after 15 epochs. The classification task involves distinguishing among four classes: glioma, meningioma, pituitary, and no tumor.

Prediction of PM2.5 concentration in urban agglomeration of China by hybrid network model

Aquaculture plays a crucial role in global food security; however, disease outbreaks remain a major constraint to sustainable production. Rapid and reliable detection of fish diseases is essential to reduce mortality, economic losses, and the misuse of antimicrobials in aquaculture systems. Conventional diagnostic approaches, such as clinical observation and bacterial culturing, are time-consuming, costly, and require specialized expertise. Recent advances in deep learning, particularly convolutional neural networks (CNNs), have shown promise in automating image-based disease detection. This study aimed to compare a lightweight three-layer CNN model with pre-trained deep learning architectures (VGG16, InceptionV3, and ResNet50) for multi-class classification of infected freshwater fish species using a balanced image dataset collected from aquaculture farms in Thailand. Images from clinically infected freshwater fish were collected during routine farm inspections across six provinces in Thailand. The dataset included 424 images of four species: Asian seabass (Lates calcarifer), red tilapia (Oreochromis sp.), snakeskin gourami (Trichopodus pectoralis), and snakeheads (Channa striata). After preprocessing, a balanced dataset of 56 images per class (totaling 224) was created. The dataset was divided into training (80%) and testing (20%) subsets. On-the-fly data augmentation techniques, such as rotation, brightness adjustment, flipping, shifting, shearing, and zooming, were applied to the training data to reduce overfitting. A lightweight, three-layer CNN model with stochastic gradient descent was used and compared with pre-trained architectures (VGG16, InceptionV3, and ResNet50). Model performance was assessed through accuracy, precision, recall, F1-score, confusion matrix analysis, and five-fold cross-validation. Among the evaluated models, InceptionV3 achieved the highest classification accuracy (56.82%), followed by VGG16 (43.18%) and the proposed CNN (38.64%), while ResNet50 performed poorly (25%). The InceptionV3 model also demonstrated higher average precision (63%), recall (57%), and F1-score (56.75%), indicating superior classification capabilities. Confusion matrix analysis revealed that InceptionV3 correctly classified 25 out of 44 test images, outperforming the proposed CNN (17 correct predictions) and VGG16 (19 correct predictions). Five-fold cross-validation further confirmed the stability and relatively better performance of the InceptionV3 model. The comparative evaluation shows that pre-trained CNN architectures, especially InceptionV3, outperform a lightweight three-layer CNN when trained on small, balanced datasets of infected fish images. Although the proposed lightweight CNN has limited accuracy, its low computational needs suggest it could be useful in resource-limited aquaculture settings. Incorporating deep learning-based image analysis into aquaculture health monitoring systems could enable quick disease triage, support timely management decisions, and promote better biosecurity and sustainable fish production. Future research should increase the dataset size, include more fish species and disease types, and test model performance across different farming environments to improve its generalizability and practical use.

Severe fever with thrombocytopenia syndrome (SFTS) is an emerging infectious disease with a high mortality rate that is often underdiagnosed due to the limitations of current laboratory testing. Timely diagnosis and early identification of severe cases are crucial to improving patient outcomes and overall survival rates. This study aimed to evaluate the efficacy of two transcripts, IFI44L and PI3, in the early differentiation between SFTS virus (SFTSV) infection and bacterial sepsis, as well as in the prompt identification of severe cases during epidemic seasons. In a prospective study conducted between 1 May 2021 and 30 September 2022, we enrolled 225 patients who presented with acute fever and thrombocytopenia at four hospitals in Shandong Province, China. The two-transcript signature provided a clear distinction between SFTS and bacterial infection, achieving an area under the receiver operating characteristic curve of 0.961 (95% confidence interval [95% CI] 0.916-0.986), outperforming C-reactive protein (0.810 [95% CI 0.738-0.870]) and procalcitonin (0.764 [95% CI 0.687-0.830]). Importantly, the relative expression of the IFI44L gene was significantly elevated in fatal SFTS cases, with an area under the curve (AUC) of 0.820 (95% CI 0.727-0.914), indicating its potential as an early prognostic marker. Additionally, IFI44L and PI3 were identified as potential biomarkers for distinguishing SFTS patients with and without invasive pulmonary aspergillosis, with AUC values of 0.817 and 0.753, respectively. Our findings demonstrate that the two-transcript signature effectively distinguishes SFTSV infection from bacterial sepsis and helps identify high-risk individuals, guiding appropriate treatment during SFTS outbreak.

The ensemble learning method is a necessary process that provides robustness and is more accurate than the single model. The snapshot ensemble convolutional neural network (CNN) has been successful and widely used in many domains, such as image classification, fault diagnosis, and plant image classification. The advantage of the snapshot ensemble CNN is that it combines the cyclic learning rate schedule in the algorithm to snap the best model in each cycle. In this research, we proposed the dropCyclic learning rate schedule, which is a step decay to decrease the learning rate value in every learning epoch. The dropCyclic can reduce the learning rate and find the new local minimum in the subsequent cycle. We evaluated the snapshot ensemble CNN method based on three learning rate schedules: cyclic cosine annealing, max-min cyclic cosine learning rate scheduler, and dropCyclic then using three backbone CNN architectures: MobileNetV2, VGG16, and VGG19. The snapshot ensemble CNN methods were tested on three aerial image datasets: UCM, AID, and EcoCropsAID. The proposed dropCyclic learning rate schedule outperformed the other learning rate schedules on the UCM dataset and obtained high accuracy on the AID and EcoCropsAID datasets. We also compared the proposed dropCyclic learning rate schedule with other existing methods. The results show that the dropCyclic method achieved higher classification accuracy compared with other existing methods.

Aims: This study aimed to explore the effect of training set diversity on the performance of deep learning models for predicting early gastric cancer (EGC) under endoscopy. Methods: Images of EGC and non-cancerous lesions under narrow-band imaging (ME-NBI) and magnifying blue laser imaging (ME-BLI) were retrospectively collected. Training set 1 was composed of 150 non-cancerous and 309 EGC ME-NBI images, training set 2 was composed of 1505 non-cancerous and 309 EGC ME-BLI images, and training set 3 was the combination of training set 1 and 2. Test set 1 was composed of 376 non-cancerous and 1052 EGC ME-NBI images, test set 2 consisted of 529 non-cancerous and 71 EGC ME-BLI images, and test set 3 was the combination of test set 1 and test set 2. Three deep learning models, convolutional neural network (CNN) 1, CNN 2 and CNN 3 (CNN 1, CNN 2 and CNN 3 were independently trained using training set 1, training set 2 and training set 3, respectively), were constructed, and their performances on each test set were respectively evaluated. One hundred and thirty-eight ME-NBI videos and 17 ME-BLI videos were further collected to evaluate and compare the performance of each model in real time. Results: On the whole, the performance of CNN 3 was the best. The accuracy (Acc), sensitivity (Sn), specificity (Sp) and area under the curve (AUC) of test set 1 in CNN 3 were 87.89% (1255/1428), 90.96% (342/376), 86.79% (913/1052) and 94.60%, respectively. The Acc, Sn, Sp and AUC of test set 2 in CNN 3 were 95% (570/600), 97.92% (518/529), 73.24% (52/71) and 90.93% respectively. The Acc, Sn, Sp and AUC of test set 3 in CNN 3 were 89.99% (1825/2028), 95.03% (860/905), 85.93% (965/1123) and 94.89%, respectively. The performance of CNN 3 was also the best in videos test set. The Acc, Sn and Sp of videos test set in CNN 3 were 91.03% (142/156), 90.58% (125/138) and 94.44% (17/18), respectively. Conclusions: The deep learning model with the most diverse training data has the best diagnostic effect.

Introduction: Every area of our lives has been devastated by the worldwide Coronavirus disease 2019 (COVID-19) epidemic. However, the development of artificial intelligence has made it possible to build advanced applications that can fulfill this level of clinical accuracy. This study aimed to create a deep learning model that can detect COVID-19 from a chest image dataset of confirmed patients treated at the provincial hospital in Aceh. Methods: Eight hundred confirmed COVID-19 patients' chest X-ray photos were gathered locally from Dr. Zainoel Abidin General Hospital, Banda Aceh. Performance was evaluated in several ways. First, the dataset was used for training and testing. Second, the data was used to train and test the model. VGG16 is a robust network adapted to an enhanced dataset constructed from a confirmed COVID-19 chest X-ray pool. To artificially produce a huge number of chest X-ray pictures, this study used data augmentation techniques such as random rotation at an angle between 10 and 10°, random noise, and horizontal flips. Results: The experimental results were encouraging: the proposed models classified chest X-ray pictures as normal or COVID-19 with an accuracy of 97.20% for Resnet50, 98.10% for InceptionV3, and 98.30% for VGG16. The results showed the outstanding performance of straightforward COVID-19 diagnosis with the classification of COVID-19 severity, such as mild, severe, and very severe. Conclusion: These made it possible to automate the X-ray image interpretation process accurately and could also be applied when materials and reverse transcription polymerase chain reaction (RT-PCR) tests are scarce.

Background Deep learning may help to improve computer-aided detection of volume (CADv) measurement of pulmonary nodules at chest CT. Purpose To determine the efficacy of a deep learning method for improving CADv for measuring the solid and ground-glass opacity (GGO) volumes of a nodule, doubling time (DT), and the change in volume at chest CT. Materials and Methods From January 2014 to December 2016, patients with pulmonary nodules at CT were retrospectively reviewed. CADv without and with a convolutional neural network (CNN) automatically determined total nodule volume change per day and DT. Area under the curves (AUCs) on a per-nodule basis and diagnostic accuracy on a per-patient basis were compared among all indexes from CADv with and without CNN for differentiating benign from malignant nodules. Results The CNN training set was 294 nodules in 217 patients, the validation set was 41 nodules in 32 validation patients, and the test set was 290 nodules in 188 patients. A total of 170 patients had 290 nodules (mean size ± standard deviation, 11 mm ± 5; range, 4-29 mm) diagnosed as 132 malignant nodules and 158 benign nodules. There were 132 solid nodules (46%), 106 part-solid nodules (36%), and 52 ground-glass nodules (18%). The test set results showed that the diagnostic performance of the CNN with CADv for total nodule volume change per day was larger than DT of CADv with CNN (AUC, 0.94 [95% confidence interval {CI}: 0.90, 0.96] vs 0.67 [95% CI: 0.60, 0.74]; P < .001) and CADv without CNN (total nodule volume change per day: AUC, 0.69 [95% CI: 0.62, 0.75]; P < .001; DT: AUC, 0.58 [95% CI: 0.51, 0.65]; P < .001). The accuracy of total nodule volume change per day of CADv with CNN was significantly higher than that of CADv without CNN (P < .001) and DT of both methods (P < .001). Conclusion Convolutional neural network is useful for improving accuracy of computer-aided detection of volume measurement and nodule differentiation capability at CT for patients with pulmonary nodules. © RSNA, 2020 Online supplemental material is available for this article.

Augmentation is a technique to increase the amount of data artificially by making more variations of the image such as changing the position of the image, changing the size of the image to changing the color of the image. In this research, the Convolutional Neural Network and K-Nearest Neighbor algorithms were used as classification methods with batik objects. The batik used is limited to 4 classes, namely Kawung, Lunglungan, Megamendung and Parang batik. The data used are 1,443 batik images. After data augmentation, 6,300 images were obtained for each technique. There are 5 augmentation techniques used, namely Random Noise, Random Rotation, Grayscale, Horizontal Flip and Vertical Flip. In this research, we succeeded in increasing the accuracy of the Convolutional Neural Network and K-Nearest Neighbor algorithms using augmentation techniques. The Convolutional Neural Network algorithm increased an average accuracy of 6% of the five augmentation techniques. Meanwhile, the K-Nearest Neighbor algorithm can increase by more than 12%. The impact of data augmentation in this study is very good, as evidenced by the increase in accuracy in both algorithms. The best accuracy is obtained when using the Vertical Flip technique, the highest accuracy is obtained in both algorithms, namely 96.92% and 81.36% on the CNN and K-NN algorithms.

The increasing incidence of carbapenem-resistant bacterial infections has become a serious public health threat. This study aimed to investigate and analyze the current regional differences in carbapenem-resistant gram-negative bacteria (CRGN) in a major Province of China, and provide suggestions for preventing hospital infections. A questionnaire survey was used to obtain the current data on CRGN from 36 hospitals in Shandong Province, China, from 2019 to 2020. The association between the detection rates and discovery rates of CRGN and the use of antibacterial drugs was analyzed using Spearman's correlation coefficient. In addition, we compared the detection rates of CRGN and antibacterial drugs using hospitals categorized according to different levels and economic areas using the Kruskal-Wallis test. The average detection rates of CRGN across the 36 hospitals varied from 1.91% to 66.04%. The discovery rate of carbapenem-resistant Enterobacteriaceae (CRE) and carbapenem-resistant Acinetobacter baumannii (CRAB) remained below 5‰, and that of carbapenem-resistant Pseudomonas aeruginosa (CRPA) was below 10‰. Except for CRAB, the correlations between the detection rate and antimicrobial drug use intensity and carbapenem drug use percentage were 0.11-0.29 and 0.31-0.47, respectively. Carbapenem drug use was higher in the provincial hospital group than in the prefecture-level hospitals (P < 0.05), and that in the high-economic regional hospital group was higher than in the low-economic regional hospital group (P < 0.05). The detection and discovery rates of CRE were low, and those of CRAB were high in Shandong Province. Larger hospitals have higher carbapenem drug use. These results can be used as a reference for preventing CRGN infections in developing countries and provide a basis for regional carbapenem resistance prevention and control strategies.

Lung cancer (LC) screening can be optimized using individuals' estimated risks of having a detectable lung tumor, as well as of mortality risk by competing causes, to guide decisions on screening eligibility, ideal screening intervals and stopping ages. Besides age, sex and smoking history, blood-based biomarkers may be used to improve the assessment of LC risk and risk of mortality by competing causes. In the German randomized Lung Screening Intervention Trial (LUSI), we measured growth/differentiation factor-15 (GDF-15), interleukin-6 (IL-6), C-reactive protein (CRP) and N-terminal pro-brain natriuretic protein (NT-proBNP), in blood serum samples collected at start of the trial. Participants in the computed tomography (CT)-screening arm also had a pulmonary function test. Regression models were used to examine these markers as predictors for impaired lung function, LC risk and mortality due to LC or other causes, independently of age, sex and smoking history. Our models showed increases in LC risk among participants with elevated serum levels of GDF-15 [odds ratio (OR)Q4-Q1 =2.47, 95% confidence interval (CI): 1.49-4.26], IL-6 [ORQ4-Q1 =2.36 (1.43-4.00)] and CRP [ORQ4-Q1 =1.81 (1.08-2.75)]. Likewise, proportional hazards models showed increased risks for LC-related mortality, hazard ratio (HR)Q4-Q1 of 4.63 (95% CI: 2.13-10.07) for GDF-15, 3.56 (1.72-7.37) for IL-6 and 2.34 (1.24-4.39) for CRP. All four markers were associated with increased risk of mortality by causes other than LC, with strongest associations for GDF-15 [HRQ4-Q1 =3.04 (2.09-4.43)] and IL-6 [HRQ4-Q1 =2.98 (2.08-4.28)]. Significant associations were also observed between IL-6, CRP, GDF-15 and impaired pulmonary function [chronic obstructive pulmonary disease (COPD), preserved ratio impaired spirometry (PRISm)]. Multi-marker models identified GDF-15 and IL-6 as joint risk predictors for risk of LC diagnosis, without further discrimination by CRP or NT-proBNP. A model based on age, sex, smoking-related variables, GFD-15 and IL-6 provided moderately strong discrimination for prediction of LC diagnoses within 9 years after blood sampling [area under the curve (AUC) =74.3% (57.3-90.2%)], compared to 67.0% (49.3-84.8%) for a model without biomarkers. For mortality by competing causes, a model including biomarkers resulted in an AUC of 76.2% (66.6-85.3%)], compared to 70.0% (60.9-77.9%) a model including age, sex and smoking variables. Serum GDF-15 and IL-6 may be useful indicators for estimating risks for LC and competing mortality among long-term smokers participating in LC screening, to optimize LC screening strategies.

Rationale: A 4-protein biomarker panel (4MP) can improve estimation of lung cancer risk and identify individuals who may benefit most from lung cancer screening. In the current study, we evaluate the performance of the 4MP for risk determination of lung cancer in individuals undergoing lung cancer screening. Methods: We obtained two cohorts of samples from the National Lung Screening Trial (NLST). The first cohort consisted of 675 samples from individuals with CT scans without suspicious findings, including 135 eventually diagnosed with lung cancer and 540 matched control samples. The second cohort consisted of 715 samples from individuals with screen-detected pulmonary nodules, including 143 diagnosed with cancer and 572 matched controls. We additionally obtained plasma from 27 cases lung cancer cases and 1308 non-cases with screen-detected pulmonary nodules that were drawn from the University of British Columbia site of the International Lung Screening Trial (ILST). The 4MP was measured using a multiplex bead-based immunoassay using coefficients fixed from a previously developed logistic regression model. Performance was evaluated using receiver operating characteristic analysis, including computing the area under the curve (AUCs) as well as a net reclassification index (NRI) to estimate how well the 4MP improved existing risk prediction models such as the Brock nodule risk model. Results: In the NLST cohort with negative CTs, for those individuals eventually diagnosed with stage II or higher lung cancer, the 4MP showed an AUC of 0.67 (95% CI 0.59-0.74). For all those with advanced (stage III and higher), AUC was 0.71 (95% CI 0.61-0.81). For individuals diagnosed within 1 year of blood draw, the AUC was 0.71 (95% CI 0.51-0.91). In those with indeterminate nodules, the 4MP showed similar performance in both the NLST and ILST samples, with an AUC of 0.64 (95% CI 0.59-0.70) in those eventually diagnosed with stage II or higher lung cancer from the NLST. We further evaluated the performance of the 4MP to reclassify pulmonary nodules, roughly based on Lung-RADS risk category. This model was developed in the ILST samples and validated in the NLST samples. The 4MP, added to the Brock model, showed an NRI of 0.26 compared to the Brock model alone in the validation set. Conclusions: The 4MP may be a useful adjunct to screening, especially in identifying those who will develop more advanced stage disease in the following year. This could help to identify individuals who may benefit from closer clinical follow-up.

Most Facial Expression Recognition (FER) systems rely on machine learning approaches that require large databases for an effective training. As these are not easily available, a good solution is to augment the databases with appropriate data augmentation (DA) techniques, which are typically based on either geometric transformation or oversampling augmentations (e.g., generative adversarial networks (GANs)). However, it is not always easy to understand which DA technique may be more convenient for FER systems because most state-of-the-art experiments use different settings which makes the impact of DA techniques not comparable. To advance in this respect, in this paper, we evaluate and compare the impact of using well-established DA techniques on the emotion recognition accuracy of a FER system based on the well-known VGG16 convolutional neural network (CNN). In particular, we consider both geometric transformations and GAN to increase the amount of training images. We performed cross-database evaluations: training with the "augmented" KDEF database and testing with two different databases (CK+ and ExpW). The best results were obtained combining horizontal reflection, translation and GAN, bringing an accuracy increase of approximately 30%. This outperforms alternative approaches, except for the one technique that could however rely on a quite bigger database.

Background and goalNoninvasive prediction of isocitrate dehydrogenase (IDH) mutation status in glioma guides surgical strategies and individualized management. We explored the capability on preoperatively identifying IDH status of combining a convolutional neural network (CNN) and a novel imaging modality, ultra-high field 7.0 Tesla (T) chemical exchange saturation transfer (CEST) imaging.MethodWe enrolled 84 glioma patients of different tumor grades in this retrospective study. Amide proton transfer CEST and structural Magnetic Resonance (MR) imaging at 7T were performed preoperatively, and the tumor regions are manually segmented, leading to the “annotation” maps that offers the location and shape information of the tumors. The tumor region slices in CEST and T1 images were further cropped out as samples and combined with the annotation maps, which were inputted to a 2D CNN model for generating IDH predictions. Further comparison analysis to radiomics-based prediction methods was performed to demonstrate the crucial role of CNN for predicting IDH based on CEST and T1 images.ResultsA fivefold cross-validation was performed on the 84 patients and 4090 slices. We observed a model based on only CEST achieved accuracy of 74.01% ± 1.15%, and the area under the curve (AUC) of 0.8022 ± 0.0147. When using T1 image only, the prediction performances dropped to accuracy of 72.52% ± 1.12% and AUC of 0.7904 ± 0.0214, which indicates no superiority of CEST over T1. However, when we combined CEST and T1 together with the annotation maps, the performances of the CNN model were further boosted to accuracy of 82.94% ± 1.23% and AUC of 0.8868 ± 0.0055, suggesting the importance of a joint analysis of CEST and T1. Finally, using the same inputs, the CNN-based predictions achieved significantly improved performances above those from radiomics-based predictions (logistic regression and support vector machine) by 10% to 20% in all metrics.Conclusion7T CEST and structural MRI jointly offer improved sensitivity and specificity of preoperative non-invasive imaging for the diagnosis of IDH mutation status. As the first study of CNN model on imaging acquired at ultra-high field MR, our results could demonstrate the potential of combining ultra-high-field CEST and CNN for facilitating decision-making in clinical practice. However, due to the limited cases and B1 inhomogeneities, the accuracy of this model will be improved in our further study.