Majid Ziaratban

Obstructive Sleep Apnea (OSA) is a prevalent and underdiagnosed sleep disorder that can lead to serious cardiovascular and cognitive complications if left untreated. This study presents a novel deep learning architecture based on convolution, LSTM, short‐time Fourier transform (STFT), attention, and transformer modules for automated OSA detection using single‐lead electrocardiogram (ECG) signals, aiming to improve diagnostic accuracy. The proposed model integrates six parallel branches for feature extraction, combining convolutional layers, recurrent units, STFT, and residual connections to capture multiscale temporal and frequency‐domain patterns. A three‐path feature refinement that incorporates sequential, convolutional, and transformer encoder is considered as the second part of the model. Channel attention‐based feature fusion modules are employed in the first and second parts to enhance feature relevance and suppress noise. Experimental evaluations on the PhysioNet Apnea‐ECG dataset demonstrate that the proposed model achieves superior segment‐level classification performance with 91.97% accuracy, 91.28% sensitivity, and 92.41% specificity. These findings suggest that the proposed method offers a robust, and scalable solution. Regarding the small number of parameters of the model, it can potentially be considered for real‐time and wearable‐based OSA monitoring applications. All codes and the trained model are released at https://github.com/mziaratban/OSA .

Background: Coronary artery disease (CAD) is recognized as an inflammatory condition and remains a leading cause of morbidity and mortality worldwide. Cardiovascular disease (CVD), more broadly, is a major contributor to global death and disability. This study aimed to compare the diagnostic performance of various noninvasive techniques for detecting CAD. Methods: A cross‐sectional study was conducted involving 758 participants, including 508 patients diagnosed with CAD and 250 without the disease. The diagnostic performance of two machine learning models—artificial neural networks (ANNs) and support vector machines (SVMs)—was evaluated. The classification models were assessed using receiver operating characteristic (ROC) curves, sensitivity, specificity, and overall accuracy. Results: The study included a total of 758 participants. Among them, 250 individuals (33.6% male and 66.4% female) were diagnosed as non‐CAD cases, while 508 participants (64% male, 36% female) were identified as having CAD. The area under the ROC curve (AUC) for CAD prediction was 0.752 (95% CI: 0.682–0.823) using the ANN model and 0.793 (95% CI: 0.733–0.853) using the SVM model. A statistically significant difference was observed between the performance of the two models in predicting CAD ( p = 0.03), with the SVM model demonstrating superior predictive performance (AUC = 0.793, 95% CI: 0.733–0.853). Conclusions: SVMs demonstrated superior performance compared to ANNs in predicting the risk of CAD using simple clinical predictors.

Accurate brain tumour classification from MRI is essential for computer‐aided diagnosis and treatment planning. This paper proposes TCAFNet, a lightweight deep learning framework that integrates multi‐scale attention mechanisms and transformer‐based refinement to enhance both local feature discrimination and global contextual reasoning. The proposed model is built upon a pre‐trained EfficientNetB0 backbone and employs an early semantic fusion strategy, in which deep features are injected into intermediate layers to guide feature learning from early stages. Cross‐scale dependencies are explicitly modelled using pairwise cross‐attention modules operating on multi‐level feature maps. Each attention‐enhanced representation is further refined using a customized adaptive convolutional block attention module and an adaptive feature pyramid attention block. The refined features are then integrated through an attention‐based fusion mechanism and processed by a lightweight Transformer block to capture long‐range spatial relationships. Experimental results show that the proposed model achieves an accuracy of 99.35% on the test set of the Figshare dataset, and demonstrates strong cross‐dataset robustness with accuracies of 99.29% and 99.67% on the test sets of Nickparvar and Br35H datasets, respectively. These results demonstrate the effectiveness of combining early feature guidance, cross‐scale attention and transformer‐based modelling for robust multi‐class brain tumour classification.

Purpose: Coronary artery disease (CAD) is one of the most significant cardiovascular diseases that requires accurate angiography to diagnose. Angiography is an invasive approach involving risks like death, heart attack, and stroke. An appropriate alternative for diagnosis of the disease is to use statistical or data mining methods. The purpose of the study was to predict CAD by using discriminant analysis and compared with the logistic regression.Materials and methods: This cross-sectional study included 758 cases admitted to Fatemeh Zahra Teaching Hospital (Sari, Iran) for examination and coronary angiography for evaluation of CAD in 2019. A logistics discriminant, Quadratic Discriminant Analysis (QDA) and Linear Discriminant Analysis (LDA) model and K-Nearest Neighbor (KNN) were fitted for prognosis of CAD with the help of clinical and laboratory information of patients.Results: Out of the 758 examined cases, 250 (32.98%) cases were non-CAD and 508 (67.22%) were diagnosed with CAD disease. The results indicated that the indices of accuracy, sensitivity, specificity and area under the ROC curve (AUC) in the linear discriminant analysis (LDA) were 78.6, 81.3, 71.3, and 81.9%, respectively. The results obtained by the quadratic discriminant analysis were respectively 64.6, 88.2, 47.9, and 81%. The values of the metrics in K-nearest neighbor method were 74, 77.5, 63.7, and 82%, respectively. Finally, the logistic regression reached 77, 87.6, 55.6, and 82%, respectively for the evaluation metrics.Conclusions: The LDA method is superior to the Quadratic Discriminant Analysis (QDA), K-Nearest Neighbor (KNN) and Logistic Regression (LR) methods in differentiating CAD patients. Therefore, in addition to common non-invasive diagnostic methods, LDA technique is recommended as a predictive model with acceptable accuracy, sensitivity, and specificity for the diagnosis of CAD. However, given that the differences between the models are small, it is recommended to use each model to predict CAD disease.

Automatic accurate skin lesion segmentation systems are very helpful for timely diagnosis and treatment of skin cancers. Recently, methods based on convolutional neural networks (CNN) have presented powerful performances and good results in biomedical applications. In the proposed method, a novel structure based on Mask RCNN, a proposed CNN, and a geodesic segmentation method is presented to improve the performance of the skin lesion segmentation. Lesions are detected and segmented by the Mask R‐CNN in the first stage. A multi‐atrous full convolutional neural network (MAFCNN) is proposed to combine outputs of the Mask RCNN and the input image to present more accurate segmentation results. To modify boundary of the lesion segmented by the MAFCNN, a geodesic segmentation method is used. Some parts of the segmentation result of the proposed CNN are utilized as labeled pixels for the geodesic method. Results demonstrate that using the proposed MAFCNN in a novel structure followed by the geodesic method significantly improves the mean Jaccard value. Experiments on five well‐known skin image datasets show that the proposed method outperforms other state‐of‐the‐art methods.

One of the most popular techniques in the Ocular Artifact (OA) removal from Electroencephalogram (EEG) signals is Adaptive Filter (AF) with Recursive Least-Squares (RLS) algorithm. The low convergence rate, good tracking, low miss-adjustment, and good stability are expected capabilities of this filter, which depend highly to value of forgetting factor (0<λ<1). As, with a λ very close to one, stability of the filter is increased due to low misadjustment, but its tracking capabilities is reduced and consequently the OA will remained in the EEG signal even after filtering. To preserve stability of the AF-RLS and improve its tracking, a new configuration of two AF-RLS in the wavelet domain is proposed for applying on approximation and detail coefficients. The proposed algorithm is compared with two older AF-RLS in the time domain (AF-T) and Wavelet based AF using approximation coefficients (WAF-A). Simulation results demonstrate the effectiveness of the proposed algorithm in term of the OA removal and preserving background EEG signals. Also some performance criteria such as visual comparison in the time domain, correlation coefficient and artifact to signal ratio are employed as evidence of this achievement. The proposed algorithm can be implemented in the real-time applications due to fast processing speed.