Biocybernetics and Biomedical Engineering最新文献

Brain cancer, one of the leading causes of mortality worldwide, is caused by brain tumors. Early diagnosis of tumors and predicting their progression can help doctors to save lives. In this article, we have designed an automated approach for locating and classifying tumors from MRI images. The novelties of the research work include the following two stages: Developing an encoder-decoder type 20-Layered deep neural network (DNN) named MultiTumor Analyzer (MTA-20) with 15 down-sampling layers and 4 up-sampling layers, the segmentation is performed in the initial stage. Here, we have adhered a Leaky ReLU activation function instead of ReLU which learn a parameter with negative values that may have valuable information which is essential specifically for image segmentation. Further, a 55-layered DNN using multistage feature fusion is developed in the second stage of the work for the classification of localized tumors. The classification is performed using developed MultiTumor Analyzer (MTA-55) DNN with Softmax classifier. The efficacy of the designed network is validated using highly cited quantitative measures such as accuracy, sensitivity, specificity, dice similarity coefficient (DSC), precision, and F1-measure. It is observed that the proposed MTA-20 DNN attains the average accuracy, sensitivity, specificity, DSC, and precision of 99.2 %, 94.6 %, 99.3 %, 88 %, and 82.5 % respectively against seven state-of-the-art techniques. Also, it is found that, the proposed MTA-55 DNN provides the overall accuracy, recall, specificity, F1-measure, precision, and DSC of 99.8 %, 99.633 %, 99.844 %, 99.659 %, 99.689 %, and 99.656 % respectively as compared to thirteen state-of-the-art techniques. These results corroborate the superiority of the proposed technique.

In addressing the challenges of medical image segmentation, particularly the elusiveness of global context and limitations in leveraging both global and local context simultaneously, we present SegmentNet as a solution. Our approach involves a step-by-step implementation within the reconstructed UNet architecture, tailored to enhance segmentation performance across diverse medical imaging modalities. The first step involves the integration of multi-focus Distance-Aware Mechanisms (DaMs) within skip connections and between successive layers of the encoder in SegmentNet. This strategic placement focuses on extracting unrelated features, ensuring comprehensive consideration of global context. Following this, Local Feature Extractor Blocks (LFEBs) are introduced at the base of the network. Equipped with depthwise separable operations, standard convolutions, smoothed ReLU, and normalization transform, LFEBs target the capture of specific local image features ensuring that features overlooked by DaMs are appropriately considered. These extracted features are then passed on to the decoder portion of SegmentNet, facilitating enhanced prediction of masks thus, optimizing segmentation performance. Evaluated across diverse datasets, including Breast Ultrasound Images (BUSI), Chest X-ray images (CXRI), and Diabetic Retinal Fundus Images (DRFI), SegmentNet excels. The segmentation evaluation results in terms of accuracy, Jaccard, and specificity are respectively recorded for BUSI, CXRI, and DRFI to be (93.88 %, 98.96 %, and 99.17 %), (99.28 %, 99.58 %, and 99.83 %), and (95.77 %, 95.95 %, and 99.94 %). Thus, showing that the incorporation of DaMs and LFEBs in SegmentNet emerges as a robust solution demonstrating precise 2D medical image segmentation across various modalities. This advancement holds significant potential for diverse clinical applications, promising improved patient care.

Squamous cell carcinoma is the most common type of cancer that occurs in many organs of the human body. To detect carcinoma, pathologists observe tissue samples at multiple magnifications, which is time-consuming and prone to inter- or intra-observer variability. The key challenge for automation of squamous cell carcinoma diagnosis is to extract the features at low (100x) magnification and explain the decision-making process to healthcare professionals. The existing literature used either machine learning or deep learning models to detect squamous cell carcinoma of specific organs. In this work, we report on the implementation of an explainable diagnostic aid system for squamous cell carcinoma of any organ and present a comparative analysis with state-of-the-art models. A classifier with an ensemble feature selection technique is developed to provide an automatic diagnostic aid for distinguishing between squamous cell carcinoma positive and negative cases based on histopathological images. Moreover, explainable AI techniques such as ELI5, LIME and SHAP are introduced to machine learning model which provides feature interpretability of prediction made by the classifier. The results show that the machine learning model achieved an accuracy of 93.43% and 96.66% on public and multi-centric private datasets, respectively. The proposed CatBoost classifier achieved remarkable performance in diagnosing multi-organ squamous cell carcinoma from low magnification histopathological images, even when various illumination variations were introduced.

Sonothrombolysis induces clot breakdown using ultrasound waves to excite microbubbles. Despite the great potential, selecting optimal ultrasound (frequency and pressure) and microbubble (radius) parameters remains a challenge. To address this, a computational model was developed to investigate the bubble behaviour during sonothrombolysis. The blood and clot were assumed to be non-Newtonian and porous, respectively. The effects of ultrasound and microbubble parameters on flow-induced shear stress on the clot surface during stable and inertial cavitation were investigated. It was found that microbubble translation towards the clot and the shear stress on the clot surface during stable cavitation were significant when the bubble was about to undergo inertial cavitation. While insonation of large microbubble (radius of $1.65\mu m$) at low frequency (0.50 MHz) produced the highest shear stress during stable cavitation, selection of these parameters is not as intuitive for inertial cavitation due to the strong competing effect between jet velocity and translational distance. An increase in jet velocity is always accompanied by a decrease in the translational distance and vice versa. Therefore, a right balance between the jet velocity and the translational distance is critical to maximise the shear stress on the clot surface. A jet velocity of 303 m/s and a distance travelled of $5.12\mu m$ at an initial bubble-clot separation of $10\mu m$ produced the greatest clot surface shear stress. This is achievable by insonating a $0.55\mu m$ microbubble using 0.50 MHz and 600 kPa ultrasound.

Calcium flux has been successfully verified to play an important role in the malignant proliferation and progression of brain tumors, which can serve as an important diagnosis guide. However, clinical diagnosis based on calcium information remains challenging because of the highly complex and heterogeneous features in calcium signals. Here we propose a calcium feature-based tumor diagnosis and treatment guidance platform (CA-TDT-GP) using random forest analysis framework for the efficient prediction of complex tumor behaviors for clinical therapy guidance. Multiple important features associated with brain tumor biological malignancy were screened out through comprehensive feature importance analysis. It provided useful guidance for understanding the biological process and the selection of drugs of brain tumors. Further clinical validation confirmed the accurate prediction of tumor biological characteristics by the model, with a coefficient of determination of over 0.86 in the same cohort of patients and over 0.77 for the new cohort of patients. We further verified the clinical malignant assessment by this model, which performed a 100% prediction match with diagnosed WHO grades, indicating great potential of the platform for clinical guidance. This promising model provides a new diagnostic and therapeutic tool for brain tumor research and preclinical treatment.

This paper proposes a fundus retinal blood vessel segmentation model based on a deep convolutional network structure and biological visual feature extraction mechanism. It aims to solve the multi-scale problem of blood vessels in the fundus retinal blood vessel segmentation task in the field of medical image processing on the basis of increasing the biological interpretability of the model. First, the subject feature information of the retinal blood vessel image is obtained by using the non-subsampled Residual Bolck convolution main channel. Secondly, combined with the study of biological vision mechanisms, an information processing model of the Retina-Exogenius-Primary visual cortex (V1) ventral visual pathway was established. Gabor functions of different scales are used to simulate the structure of different levels of the visual pathway, and the scale information at different levels is integrated into the corresponding hierarchical stages of the convolutional main pathway network to enrich the information of small blood vessels and enhance the semantic information of the overall blood vessels. Finally, considering the imbalance of the ratio of vessel and nonvessel pixels, an adaptive optimization scheme using hybrid loss function weights is proposed to enhance the priority of blood vessel pixels in the calculation of the loss function. According to the experimental results on the STARE, DRIVE and CHASE_DB1 data sets, the model still achieves superior performance evaluation indicators overall compared with the existing optimal methods in the fundus retinal blood vessel segmentation task. This research is of great significance to the field of medical image processing and can provide more accurate auxiliary diagnostic information for clinical diagnosis and treatment.

This study investigates the flow and hemodynamics of small perforator blood vessels that branch from the basilar artery (BA) in the brain. Using advanced imaging techniques and computational fluid dynamics (CFD) simulations, detailed 3D geometries of the perforators were acquired through barium contrast injection, micro-CT scans, and data processing. The hybrid geometry, combining micro-CT scans and mesh extraction algorithms, provided accurate vessel models. The influence of different types of finite volume on the analysis was examined, with polyhedral elements showing the most efficient ratio of the analysis time to convergence level. Additionally, the effect of boundary conditions on hemodynamic parameters was studied. Simulations using 0.0 mmHg pressure conditions at the outlets directed flow mainly through the BA, neglecting the perforator branches. In contrast, non-zero outlet pressure conditions significantly increased the flow in the perforators, leading to non-physiological flow velocities and overestimation of hemodynamic parameters. The assumption of pressure conditions of 0 mmHg at outlets was found to be valid for simple single vessel geometries, but not for more complex vascular systems. This research contributes valuable information on the complex flow patterns and hemodynamics of small perforator blood vessels in the brain and emphasizes the importance of accurately modeling geometry and boundary conditions in such studies.

The paper proposes a new model-based optimization approach to improve the clinical efficiency of compensatory insulin bolus treatment in diabetic patients, aiming to mitigate the consequences of diabetes. The most important contribution of this paper is a novel methodology for determining the optimal parameters of insulin treatment, namely the size and timing of insulin boluses, to effectively compensate for carbohydrate intake. This concept can be seen as the so-called optimal model-based bolus calculator. The presented theoretical framework deals with the problem of optimal disturbance rejection in impulsive systems by minimizing an integral quadratic cost function. The methodology considers a personalized empirical transfer function model with static gains and time constants as the only parameters assumed to be known, making the bolus calculator more straightforward to implement in clinical practice. Contrary to other techniques, the proposed methodology considers impulsive insulin administration in the form of boluses, which is more feasible than continuous infusion. In contrast to the conventional bolus calculator, the proposed algorithm allows for maximizing therapy performance by optimizing the relative time of insulin bolus administration with respect to carbohydrate intake. Another feature to highlight is that the solution of the optimization problem can be obtained analytically, hence no numerical iterative solvers are required. Additionally, the continuous-time domain approach allows for a much finer adjustments of the insulin administration timing compared to discrete-time models. The proposed approach was validated in an in-silico study, which demonstrated the importance of systematically determined insulin–carbohydrate ratio and the relative delay between disturbance and its compensation. The results showed that the proposed optimal bolus calculator outperforms the traditional suboptimal formula.

The diverse shapes and scales, complicated backgrounds, blurred boundaries, and similar appearances challenge the current organ segmentation methods in medical scene images. It is difficult to acquire satisfactory performance to directly extend the object segmentation methods in the natural scene images to the medical scene images. In this paper, we propose a discriminant feature pyramid (DFPNet) network for organ segmentation in the original medical images, which consists of two sub-networks: the feature steered network and the border network. To be specific, the feature steered network takes a top-down step-wise manner to extract abundant context information, which is conducive to suppressing the cluttered background and perceiving the scale variation of objects. The border network utilizes a bottom-up step-wise manner to optimize the boundary feature map, which aims at distinguishing adjacent edge features with similar appearances but diverse labels. A series of experiments were conducted on three publicly available medical datasets ( i.e., LUNA 16, RIM-ONE-R1, and VNC datasets) to evaluate the validity and generalization of the proposed DFPNet. Experimental results indicate that our network achieves superior performance in terms of the receiver operating characteristic (ROC) curve, F-Score, Jaccard index, and Hausdorff distance. The code will be available at: https://github.com/Xie-Xiwang/DFPNet.

The recognition task of visual stimuli based on EEG (Electroencephalogram) has become a major and important topic in the field of Brain–Computer Interfaces (BCI) research. Although the underlying spatial features of EEG can effectively represent visual stimulus information, it still remains a highly challenging task to explore the local–global information of the underlying EEG to achieve better decoding performance. Therefore, in this paper we propose a deep learning architecture called Linear-Attention-combined Convolutional Neural Network (LACNN) for visual stimuli EEG-based classification task. The proposed architecture combines the modules of Convolutional Neural Networks (CNN) and Linear Attention, effectively extracting local and global features of EEG for decoding while maintaining low computational complexity and model parameters. We conducted extensive experiments on a public EEG dataset from the Stanford Digital Repository. The experimental results demonstrate that LACNN achieves an average decoding accuracy of 54.13% and 29.83% in 6-category and 72-exemplar classification tasks respectively, outperforming the state-of-the-art methods, which indicates that our method can effectively decode visual stimuli from EEG. Further analysis of LACNN shows that the Linear Attention module improves the separability between different category features and localizes key brain region information that aligns with the paradigm principles.