Medical & Biological Engineering & Computing最新文献

Schwannomas (SCH) and meningiomas (MEN), the two most common primary spinal cord tumors, present a clinical diagnostic challenge due to their overlapping clinical and radiological manifestations. To address this, we developed a deep learning-based object detection model for automated detection of these tumors using magnetic resonance imaging (MRI), which could facilitate early diagnosis and alleviate clinical decision-making burdens. Our study retrospectively analyzed MRI scans from 103 pathologically confirmed SCH and MEN cases at a local hospital (July 2015-August 2024). First, we took YOLOv8n as the baseline model, introduced selective kernel fusion (SKFusion) module to replace the feature fusion layer of the original neck part, added recursive gated convolution (gnConv), and then trained the improved feature fusion model (YOLOv8n-SKNeck). The proposed model achieved notable performance metrics: 91.20% mean accuracy, 90.92% mean recall, and 91.03% mean F1-score for SCH/MEN detection. These results demonstrate that our optimized deep learning framework can effectively automate the detection and differential diagnosis of spinal SCH and MEN through MRI analysis. Thus, the novel method holds significant potential for advancing computer-aided diagnosis and facilitating innovative applications in future clinical practice.

During puncture procedures, respiratory motion can cause significant displacement of lesions. Fast and accurate estimation of pulmonary respiratory motion can provide valuable guidance during surgery. However, the large deformation of fine lung textures and the complex motion of internal structures such as airways and blood vessels pose significant challenges for motion estimation. In this study, we propose a multi-resolution parallel network architecture based on neural ordinary differential equations (neural ODE). By incorporating neural ODE, our method explicitly models the temporal continuity of 4DCT data, addressing the issue of unrealistic deformations in lung motion estimation and producing transformations that better align with the physiological patterns of respiratory motion. Furthermore, we introduce a multi-resolution parallel structure to recursively refine lung features. This enhances the network's feature representation and prediction capabilities, thereby improving registration accuracy. We conducted both qualitative and quantitative experiments on the TCIA and DirLab datasets, demonstrating that the proposed method outperforms other deep learning approaches and achieves consistently high performance across all respiratory phases.

Diabetic retinopathy (DR), a common diabetic complication, stands as a primary cause of retinal blindness. Rapid automatic DR grading is a crucial approach for prevention. Although deep learning has shown promising results in automated DR grading, its deployment in clinical settings remains challenging due to variations in imaging devices, lighting conditions and other conditions in different hospitals. This study explores the problem of decreased generalization performance in DR grading, stemming from variations in the distributions of the source and target domains, namely addressing the domain generalization (DG) problem in DR grading. To tackle this challenge, we propose a novel Fourier-based domain generalization framework for fundus images, which consists of three key innovation elements: (1) Fourier spectrum enhancement: a Fourier-based image enhancement technique preserves critical high-frequency features by leveraging phase information, significantly improving cross-domain robustness; (2) Collaborative teacher-student knowledge distillation: a dual-network learning mechanism enhances generalization by distilling high-level semantic features from a teacher model to a student model; (3) Feature fusion: a feature fusion module effectively differentiates intra-class and inter-class features, thereby further enhancing classification performance. Extensive evaluation of our framework on six clinically realistic DR datasets demonstrates superior generalization performance compared to existing methods. Furthermore, this study reveals the critical role of Fourier phase information and high-level semantic features in improving generalization, bridging an important research gap in DR grading.

In patients with developmental dysplasia of the hip (DDH), complications such as head dislocation and cup instability after total hip arthroplasty (THA) are often caused by poor cup positioning. However, no planning algorithm currently considers both the cup bony coverage rate (BCR) and the hip impingement-free range of motion (IFROM) simultaneously due to the high variability in cup position and the difficulty in parameterizing IFROM after THA. This study proposed an algorithm to efficiently calculate the BCR and IFROM for all potential cup positions, filtering out the cup's optimal implant position. The IFROM evaluation coefficients (EC_IFROM) were calculated based on the maximum IFROM for each cup position. Cup positions meeting the BCR requirement and ranking in the top 10% of EC_IFROM were selected as the optimal implant positions. The algorithm was tested on three bone models with varying degrees of DDH to compare BCR and optimal implant position. The algorithm developed in this study is versatile, reliable, and efficient. Calculations in three patients showed that (1) the roof BCR is usually greater than the 3D BCR, and (2) the EC_IFROM makes it easier to select the optimal implant position of the cup. This method can calculate the BCR and IFROM for all cup positions, providing surgeons with a reference to determine the optimal cup position and whether augmentation is needed to improve cup stability in patients with DDH, thereby reducing the incidence of impingement after THA.

Elevated intracranial pressure (ICP) is recognized as a biomarker of secondary brain injury, with a significant linear correlation observed between optic nerve sheath diameter (ONSD) and ICP. Frequent monitoring of ONSD could effectively support dynamic evaluation of ICP. However, ONSD measurement is heavily reliant on the operator's experience and skill, particularly in manually selecting the optimal frame from ultrasound sequences and measuring ONSD. This paper presents a novel method to automatically identify the optimal frame from video sequences for ONSD measurement by employing the kernel correlation filter (KCF) tracking algorithm and simple linear iterative clustering (SLIC) segmentation algorithm. The optic nerve sheath is mapped and measured using a Gaussian mixture model (GMM) combined with a KL divergence-based method. When compared with the average measurements of two expert clinicians, the proposed method achieved a mean error, mean squared deviation, and intraclass correlation coefficient (ICC) of 0.04, 0.054, and 0.782, respectively. The findings suggest that this method provides highly accurate automated ONSD measurements, showing potential for clinical application.

Extracorporeal membrane oxygenation (ECMO) is a life-support modality that supports cardiopulmonary function in critically ill patients. Veno-arterial (VA)-ECMO, which reinfuses blood through the femoral artery, can cause upper body hypoxemia due to uneven oxygen distribution within the aorta. In this study, computational haemodynamic simulations were performed using a patient-specific aorta geometry to simulate blood flow and quantify oxygen saturation (SO₂) levels in the arch branches under varying ECMO support levels. A single-phase flow model coupled with oxygen transport was employed and compared to the multiphase flow model used in previous studies. Simulation results were analysed to evaluate the locations of watershed zones formed by the interplay of native cardiac and ECMO flows and the corresponding oxygen levels. Our findings demonstrate that the single-phase model predicted IA SO₂ ranging from 71.1% to 94.2%, whereas the multiphase model predicted 70.0-86.7%, indicating an underestimation of oxygen saturation in the aortic arch branches. Additionally, lower ECMO support levels shift the watershed region distally, reducing oxygen delivery to the arch branches. This study highlights the potential of computational haemodynamic simulations for assessing oxygen transport and haemodynamic behaviour in patients undergoing VA-ECMO. The approach provides insights to support clinical decision-making and improve personalised treatment outcomes.

Cerebrovascular regulation, driven by mechanisms such as cerebral autoregulation and the Cushing's reflex, plays a critical role in maintaining cerebral blood flow (CBF) adequate despite changes in arterial pressure (AP), since a dampening of CBF can lead to serious brain pathologies. This study investigates the causal and self-predictable dynamics of cerebrovascular interactions in patients undergoing coronary artery bypass graft surgery, before and after propofol general anaesthesia. The dynamics of the pressure-to-flow and flow-to-pressure links between mean arterial pressure (MAP) and mean cerebral blood velocity (MCBv) is assessed using time-domain and frequency-domain measures of Granger Causality (GC) and Granger Autonomy (GA). The results indicate that while time-domain indices remain stable, frequency-domain measures reveal variations in the very-low-frequency, low-frequency, and high-frequency (HF) bands. The increased spectral GC in the HF band may be related to the effect of mechanical ventilation during anaesthesia. Additionally, a reduction in self-dependency of MCBv in the HF band reflects weakened internal regulatory mechanisms post-anaesthesia. In conclusion, propofol-induced suppression of sympathetic control and the effects of mechanical respiration increase the dependence of cerebral blood flow on arterial pressure in specific bands of cerebrovascular interest. These findings underscore the importance of frequency-domain analysis in detecting subtle cerebrovascular dynamics that time-domain measures may overlook.

In this paper, we extend the SonoNet architecture to capture spatio-temporal information from ultra-sound (US) sequences. More specifically, we propose 3D-SonoNet32 - which lifts 2D convolutions to 3D - and to an efficient (2+1)D variant - to keep the computational cost under control while preserving the benefits of the spatio-temporal model. We investigate the potential of these architectures on a scan-plane detection problem and discuss how these methodologies can be beneficial for AI-driven online "scan assistants", to enhance the quality and reproducibility of the evaluation and ultimately support the clinicians in the US examination. Our main contributions are (i) the design of novel Space-Time SonoNet architectures for analysing US video sequences, (ii) an in depth experimental analysis to show the benefit of using space-time models with respect to purely spatial ones, and to discuss the potential improvements gained by exploiting domain-specific properties like temporal coherence and prior knowledge of the ongoing scan. Overall, we show that the proposed models are specifically designed to be computationally lightweight, but also competitive in performance, making them suitable for real-time deployment on portable US devices.