Myocardial infarction (MI) is a life-threatening cardiac condition that necessitates early diagnosis and accurate severity assessment for effective clinical management. Existing automated techniques mostly focus on differentiating MI from healthy controls (HCs), limiting their clinical utility by overlooking severity staging and quantitative risk evaluation. To address this gap, we propose a novel attentive hybridmultiscale fusion network (AHMFN) for simultaneously classifying MI severity stages and estimating quantitative risk score (RS) levels. The network integrates dual-stream CNN with both fixed and variable receptive fields for multiscale spatial feature extraction, utilizes multihead attention for improved feature weighting, and a BiLSTM module to capture temporal dependencies in electrocardiogram (ECG) data. The proposed RS quantitatively measures diagnostic certainty based on cardiac axis deviation (CAD). Comprehensive evaluations of the PTB-XL dataset exhibit an accuracy of 99.82% for MI severity stages and 95.2% for RS levels, while external validation on the PTBDB and STAFF III datasets attained an accuracy of 99.5%, hence affirming generalizability. RS further exhibits robust clinical interpretability, with a Pearson correlation of 0.85 to CAD angles. Despite its advanced design, AHMFN ensures deployment feasibility with <= 89 ms/step inference latency and a compact similar to 2.15 MB model size, making it suitable for real-time wearable applications. This balance of efficiency and accuracy enables interpretable early-stage diagnosis and timely intervention, while outperforming existing methods.

Accurate and effective cardiovascular disease (CVD) diagnosis is particularly difficult in telemedicine and resource-constrained environments due to traditional multi-lead ECG devices' high computational and operational expenses. We propose a computationally effective graph-based approach to the automated detection of CVD from reduced-lead {I, II} electrocardiogram. The approach formulates lead relationships as a dynamic graph $G=(V, E)$ whose nodes $V=\lbrace \text{I}, \text{II}\rbrace$ correspond to leads and whose edge weights $w_{ij}(t) \in E$ capture time-varying cardiac axis deviation angles $\theta (t)$ in the frontal plane. Three statistical features are obtained mean angle $\mu _\theta$, angular variance $\sigma _\theta ^{2}$, and lead correlation coefficient $\rho _{\text{I, II}}$. Experimental testing on PTB-XL and PTB datasets establishes state-of-the-art performance at 89.2% and 84.1% accuracy, respectively, without redundant computations native to multi-lead ECG. The approach ensures clinical-grade accuracy with $O(1)$ feature extraction complexity, providing an optimal trade-off between accuracy and computational efficiency for resource-constrained wearable ECG sensors and tele-ECG applications.

Age-related macular degeneration (AMD) is a retinal disease that can impair the central vision permanently. Accurate delineation of the retinal pigment epithelium (RPE) and Bruch’s membrane (BM) in optical coherence tomography (OCT) B-scans is crucial for diagnosing and monitoring AMD. While automated segmentation methods exist for early AMD stages, late-stage AMD remains a challenging area due to the pronounced disruption of the RPE and BM. To ensure spatial contiguity in the boundary delineation of RPE and BM, both the global and local contextual information must be learned. In this context, we propose a generative adversarial network (GAN) to segment these significant retinal interfaces in OCT B-scans from AMD patients. A UNet++ model with its deep supervision is trained using a hybrid loss function combining adversarial loss and multi-class cross-entropy (CE) segmentation loss. The CE loss learns the local features by optimizing the per-pixel accuracy, while the adversarial loss captures a broader context by learning overall layer label statistics. This loss combination allows the model to capture fine details in the ordered retinal layer structure and guide layer boundaries along discontinuities in the RPE and BM in severe AMD cases. Additionally, a graph search algorithm refines boundary delineations from predicted segmentation maps. The model’s effectiveness is validated on the DUEIA and AROI datasets, which include OCT B-scans from both AMD-affected and healthy individuals. The proposed approach achieves Mean Absolute Errors (MAE) of 0.45 and 1.19 on the respective datasets, demonstrating its capability to handle boundary segmentation in severe AMD cases.

Brain shift is an undesired phenomenon during brain surgery, which is effectively managed by image registration between preoperative and intraoperative brain images. In ultrasound-guided brain surgery, the similarity match between registering image pairs must be sufficiently large, which cannot be obtained due to the poor resolution of the ultrasound (US) image. Segmentation-based approaches for registration were found useful in the literature. Many regions in the preoperative brain ultrasound (pBUS) image do not match the corresponding regions in the intraoperative brain ultrasound (iBUS) image well due to speckle and other artifacts. In an earlier work, we presented a level-set-based segmentation method of different anatomical regions of brain ultrasound images to separate five different common prominent anatomical regions from pBUS and iBUS images. In this article, we use the said segmentation method as prework for the registration method. Due to the low similarity score of US images, segmentation of common anatomical regions is performed from both pBUS and iBUS images. An initial solution is found from the user-provided box marks before segmentation. The proposed registration model is rigid, and both traditional and heuristic-based optimization techniques are applied. The outcomes are compared with the state-of-the-art methods in this field.

Conduction disturbances (CD) are cardiac abnormalities characterised by disruptions in the cardiac electrical conduction system. Accurately identifying the risk score (RS) for CD in electrocardiogram (ECG) signals is essential for early diagnosis and effective treatment of cardiac disease. The RS offers a quantitative assessment of diagnostic certainty by analysing cardiac axis deviation (CAD) and ECG morphological alterations. Current methods mostly focus on the diagnosis of cardiovascular diseases (CVDs), frequently overlooking the intricate ECG features vital for thorough risk assessment. This paper proposes a novel application for CD classification and associated risk identification using a hybrid Convolutional Long Short-Term Memory (CLSTM) network. The model integrates convolutional layers with various filter sizes to capture multiscale spatial features and LSTM layers to address temporal dependencies. The network concludes with a fully connected layer utilising softmax activation to classify RS over four levels and CD into seven categories. The model was trained using the PTB-XL dataset, achieving an overall accuracy and F1-score of 85.8% and 81% for CD classification and 99.6% and 96% for RS prediction, respectively. By demonstrating the model's proficiency in capturing intricate ECG features, these findings highlight its potential to enhance cardiac diagnosis.

Automating CVD diagnosis is critical for effective cardiac monitoring; however, challenges arise from subtle structural variations in ECG data under different heart conditions. Current methodologies mostly emphasise extracting lead-specific morphological characteristics from multi-lead ECG data, often neglecting the spatial relationships across leads and temporal dependencies of ECG signals. This paper presents a novel temporal convolutional bidirectional LSTM (CBiLSTM) method for the classification of five cardiac ailments. The model includes BiLSTM for capturing past and future bidirectional temporal dependencies and convolutional layers to collect spatial information. The model was trained using the PTB-XL dataset, yielding an overall accuracy and F1-score of 75.8% and 75%, respectively. The findings demonstrate that the proposed strategy outperforms current methods in the classification of CVD and enhances the early detection of cardiac abnormalities.

Synthesis of a 12-lead electrocardiogram from a reduced lead set has previously been extensively studied in order to meet patient comfort, minimise complexity, and enable telemonitoring. Traditional methods relied solely on the inter-lead correlation between the standard twelve leads for learning the models. The 12-lead ECG possesses not only inter-lead correlation but also intra-lead correlation. Learning a model that can exploit this spatio-temporal information in the ECG could generate lead signals while preserving important diagnostic information. The proposed approach takes leverage of the enhanced inter-lead correlation of the ECG signal in the wavelet domain. Long-short-term memory (LSTM) networks, which have emerged as a powerful tool for sequential data mining, are a type of recurrent neural network architecture with an inherent capability to capture the spatiotemporal information of the heart signal. This work proposes the deep learning architecture that utilizes the discrete wavelet transform and the LSTM to reconstruct a generic 12-lead ECG from a reduced lead set. The experimental results are evaluated using different diagnostic measures and similarity metrics. The proposed framework is well founded, and accurate reconstruction is possible as it can capture clinically significant features and provides a robust solution against noise.

Retinal edema caused by fluid buildups has links to vision threatening retinal diseases. Presently, spectral-domain optical coherence tomography (SD-OCT) is the imaging method of choice for evaluating retinal health and tracking treatment progress. Automated methods for fluid segmentation in retinal OCT B-scans are critical for ocular disease diagnosis. The segmentation task in retinal OCT B-scans with pathological manifestations is challenging as the retinal layered structure varies significantly with the severity of the underlying disease condition. In this paper, we present a Contextual Self-Attention based U-Net (CoSAUNet) architecture for fluid segmentation in pathological OCT B-scans. The proposed model replaces the basic convolutional layers in the standard U-Net architecture with contextual self-attention layers. The contextual self-attention mechanism integrates contextual information mining among neighboring keys and conventional self-attention into a unified architecture to learn the long-range feature dependencies. The robustness of the proposed CoSAUNet model is evaluated through a 3-fold cross-validation setup on the RETOUCH and the AROI datasets, yielding average Dice scores (DS) of 88.16% and 76.67%, respectively. Furthermore, the CoSAUNet performs comparably to the state-of-the-art methods while using only half of the parameters.

The recent advancement in deep learning (DL) for automatic modulation classification (AMC) of wireless signals has encouraged numerous possible applications on resource-constrained edge devices. However, developing optimized DL models suitable for edge applications of wireless communications is yet to be studied in depth. In this work, we perform a thorough investigation of optimized convolutional neural networks (CNNs) developed for AMC using the three most commonly used model optimization techniques: a) pruning, b) quantization, and c) knowledge distillation. Furthermore, we have proposed optimized models with the combinations of these techniques to fuse the complementary optimization benefits. The performances of all the proposed methods are evaluated in terms of sparsity, storage compression for network parameters, and the effect on classification accuracy with a reduction in parameters. The experimental results show that the proposed individual and combined optimization techniques are highly effective for developing models with significantly less complexity while maintaining or even improving classification performance compared to the benchmark CNNs.

Cooperative spectrum sensing among cognitive radio (CR) enabled smart devices improves the sensing performance in deep fading environments. However, it is prone to attacks by malicious users. These attacks become more severe when they are launched in collusion. In this paper, a collision penalty-based attack prevention method is proposed to protect honest CR enabled smart devices from individual and collusion attacks. First, an optimal decision fusion rule which maximizes the throughput of each CR enabled smart device is proposed. Then, we identify the possible malicious strategies and prevent them with appropriate bounds on the penalty. Investigations are based on the malicious utility obtained by the attackers after playing malicious strategies, and the utility reduction capability of an attack prevention mechanism. Simulations demonstrate that the proposed mechanism can achieve up to 700% of malicious utility reduction as opposed to 150% and 100% in the cases of existing Moral Hazard Principal Agent (MHPA) and friendly jammer attack prevention schemes. A game theoretic analysis of the proposed method shows that a unique Nash Equilibrium is achieved when all the CR enabled smart devices are honest.

In this research paper we propose a novel technique is to predict the Confidence Score of Myocardial Infarction from multilead electrocardiogram (ECG) signals. Classifying the confidence score (CS) of myocardial infarction (MI) severity provides a quantitative measure of the severity level, enabling accurate assessment. MI v/s non MI is used in automated diagnosis but CS is not used. If any patients having MI then level of disease can be captured by CS which can help cardiologist for better treatment planning. ECG morphological alterations associated with the progression of MI show pathological traits such T-wave inversions, changes in ST elevation/depression, pathological Qwaves, hypercute T-waves. Multiscale wavelet energies and eigen values of multiscale covariance matrices are used as diagnostic features for prediction of confidence score of MI. As classifiers, K-nearest neighbor (KNN), Decision Tree, Random Forest, AdaBoost and Gradient Boosting are used. The PTB-XL database used for evaluation, comprises different types of MI, healthy control (HC) subjects, and five distinct classes of confidence scores (15, 35, 50, 80, 100). The result for classification of HC and MI with confidence score evaluated accuracy & F1 score of 96.7%, 98.5% respectively on PTB-XL database using Random forest which is better than existing state of art. The overall accuracy performance of CS of MI for CS =15, CS =35, CS =80, CS =100 are are 98.2%, 63%, 63.7% and 90.5% respectively using Gradient boosting classifier. The usage of various classifiers demonstrates their effectiveness in predicting different classes of confidence scores, providing valuable insights for further analysis.

Brain ultrasound (BUS) is a plausible solution for both pre-operative and intra-operative brain imaging during neurosurgery. The registration between pre-operative and intra-operative images resolves the adverse effect of the brain shift. But this registration is challenging due to various artefacts present in the intra-operative BUS images. This paper presents a segmentation-based approach for the registration of pre- and intra-operative brain ultrasound images that reduces the effect of artefacts present in both the images. The similarity between pre-operative brain ultrasound image (pBUS) and intra-operative brain ultrasound(iBUS) images is poor because of the speckle and present artefacts. It makes registration challenging. It was observed that the hyper-echoic (HE) regions in pBUS and iBUS images get less affected by the artefacts, and these regions have considerable similarity. We apply a patch-based segmentation approach to both the pBUS and the iBUS images to separate those HE regions. Rigid registration is performed between the segmented HE regions. The logarithm of the mean squared error(MSE) captures the similarity between registering image pairs. Different heuristic optimization algorithms such as biogeography-based algorithm (BBO), simulated annealing (SA), and particle swarm optimization (PSO) were applied to minimize the logarithm of MSE. Comparing the results, we found that PSO outperforms all other algorithms in order to find the registration parameters. Experiments were conducted on the images from the BITE and the RESECT datasets. Results of the proposed method were compared with two other existing methods. Common tag-points were marked with the help of an expert radiologist for evaluation of the performance of the registration. The mean target registration error (mTRE) and structural similarity index (SSIM) were measured before and after the registration for the comparable methods. The average mTRE was computed on 43 pairs of images. The mTRE before registration was 5.87 mm that reduced to 4.8 mm by the method of Chel and Bora (2017), whereas the proposed method reduced it to 2.91 mm. The feature-based method (Machado et al. 2018) failed because of improper feature matching between the registering image pairs. Unlike other methods, the proposed method considers the effects of artefacts present in the registering pBUS and iBUS image pairs. The proposed method adopts a patch-based segmentation approach which is robust to noise and extracts similar HE regions from the registering image pairs. Registration is performed between the segmented HE regions from the pBUS and iBUS images. Performance of the proposed method was compared with another PSO-based method (Chel and Bora 2017) and a feature-based method (Machado et al. 2018). The proposed method outperformed the other two methods in reducing the mTRE and SSIM between the registering images.

With the increase in the number of documents with various types of charts available on the internet, automatic chart classification has become an essential task for various downstream applications such as chart data recovery, chart replenishment. This paper presents a comprehensive survey of the studies reported in the literature since 2001 from the perspective of the corpus, pre-processing techniques, feature extraction, and methodologies. Considering that the majority of the existing studies use small datasets with a smaller number of chart types and also reported varying performances, this paper implements and evaluates 44 different machine learning-based chart classification models. The evaluation is done over a large dataset curated locally and benchmarks the performances of these 44 different models over a common experimental framework. It also performs a comprehensive error analysis, identifying two core challenging issues (noise in the charts and confusing chart pairs) that affect the chart classification performances. Compared with the existing survey papers, this paper presents a more comprehensive review and experimental analysis.

As we delve deeper into the features learned by networks associated with various computer vision tasks, such as object detection and image classification, we find that they vary from more general to more network-specific. Feature space-based adversarial perturbation methods, which exploit these general features in the earlier layers of the surrogate network, can generate adversarial examples transferable to a target network. Image forensics networks, such as those associated with camera model identification and manipulation detection, are generally shallow and trained from scratch. Such networks are said to be devoid of general features. Recent works in image forensics using various gradient-based adversarial generation methods show the limited transferability of these adversarial examples. A stop condition added to the gradient-based generation algorithms leads to an improvement in transferability. In this work, we exploit a feature space-based perturbation method to examine the generality of the learned features of the image forensics networks and generate transferable adversarial examples. Under various knowledge scenarios, the transferability in terms of the attack success rate is task dependent. The transferability achieved in the median filtering case is significant and shows the generality of the learned features in image forensics networks. The same cannot be said for the resizing detection task, as the transferability is found to be zero under all knowledge scenarios.

Segmentation of different regions in intra-operative brain ultrasound (iBUS) images is often required for assisting the neuro-surgeon. Traditional level-set and active contour-based semi-automatic image segmentation approaches suffer from low accuracy and slow convergence. This paper presents a novel semi-automatic level-set approach for segmenting hyper-echoic (HE), hypo-echoic, and anechoic regions with minimal user intervention. Three HE regions longitudinal fissure, choroid plexus, and tumor and two anechoic regions, namely ventricle and resection cavity are segmented using a patch-based level-set approach. This method is a combination of three procedures: a) unidirectional level-set curve flow (ULSCF), b) bidirectional level-set curve flow (BLSCF) using a logarithmic patch size control, and c) cubic B-spline-based contour smoothing. The zero level-set curve is derived using a patch-based intensity thresholding method of the desired region. The imperfection on the blocky edges produced during the patch-based ULSCF, are minimized using the BLSCF step that uses a local region splitting approach. Slope and curvature discontinuities of the resulting boundaries after BLSCF are eliminated using cubic B-spline based contour smoothing. The proposed method outperforms other state-of-the-art level-set and active contour methods, and the desired result is obtained within reasonable time required for online monitoring during surgery.

With the advancement in the world of digitization, storing information in the form of scanned copies, images, etc. becomes a new normal. This new normal leads to the need for a system that can extract accurate information from the scanned documents or images with respect to every component they may have, such as textual, graphical, etc. The first step in extracting document information is to segment the document layout: divide the document into textual and non-textual regions of interest. There have been various studies over document layout segmentation, and this study observed that the majority of the existing studies face one common challenge, i.e., accurate segmentation of graphical components with sparsely clustered pixels such as flowcharts, block diagrams, etc. The study addresses it with a two-tier feedback-based framework. The first tier segments and classifies the textual and mathematical equation components, while the second tier segments and classifies the graphical regions using the feedback information from the first tier. The information provided by the first tier is the regional information of the equation and textual components to get a different copy of the original input document image in such a way that most of the foreground pixels are part of graphical regions. The proposed framework outperforms various existing studies (when evaluated against multiple data sets).

This paper presents a patient-specific approach for reconstructing the standard 12-lead ECG from a minimal lead set. The 12-lead ECG signal acquisition impediment using ten electrodes comprises ambulatory monitoring, personalized healthcare, remote healthcare, and pediatric ECG lead placement. Furthermore, concurrently processing signals from multiple electrodes enhances the intricacy as well as the cost. Synthesizing 12-lead ECG from a reduced lead set becomes a better solution. This study proposes a recurrent neural network (RNN) long short-term memory (LSTM) to synthesize standard 12-lead ECG from the three predictor leads. The four performance metrics, namely correlation coefficient (cc), root mean square error (RMSE), and wavelet energy diagnostic distortion (WEDD), are employed to evaluate the performance of the proposed method. The proposed model obtained fine reconstruction quality and achieved better performance than most of the previously established works without compromising diagnostic information.

The electrocardiogram (ECG) based biometric system has been gaining wide attention in recent years. ECG has many attractive characteristics for human identification applications, such as it is easily measured, universal, difficult to counterfeit, and offers additional information on physiological and clinical traits. In this paper, we attempt to analyse the effectiveness of the synthesised ECG signal for person identification applications. This paper also investigates the similarity of the biometric information present in the original signals and their reconstructed signals of the same record. We have proposed a novel performance measure, i.e., person identifiability (PI), to evaluate the synthesis model. An LSTM based neural network model has been employed to synthesise the ECG leads. The person identification performance for the synthesised ECG leads is obtained using the HLSTM model. The proposed framework is tested using the PTB ECG database.

Because of the rapid growth in the VLSI technology, the speed of the Internet and the easy availability of editing software, the number of near-duplicate videos (NDVs) produced, edited or viewed has increased manifold. This has made the effective database management, and the video copyright protection is a challenging problem. To identify the NDVs in the database, a perceptual video hashing scheme for the content-based retrieval of NDVs is proposed in this paper. The perceptual video hashing algorithms used to design the near-duplicate video retrieval (NDVR) application are based on (1) the Achlioptas’s random matrix (ARM), (2) the temporal wavelet transform (TWT) and the ARM, (3) the Tucker decomposition and (4) the parallel factor (PARAFAC) decomposition. The performance of the NDVR application is evaluated using the average precision–recall curves on three video databases created from REEFVID, XIPH, OPEN-VIDEO and TRECVID data sets using various content-preserving operations and malicious modifications. The experimental results on a moderate-size video databases show that the Tucker decomposition-based video hashing algorithm performs better in retrieving NDVs.