Introduction Delineating infiltrative peritumoral edema (PTE) in gliomas remains challenging due to acquisition variability and the lack of histopathological validation. We systematically assessed artificial intelligence (AI) and radiomics models for peritumoral characterization.Methods Following PRISMA, we searched five databases on July 18, 2025, for peritumoral AI/radiomics studies predicting tumor infiltration, recurrence, or molecular status. Quality was evaluated using Prediction model Risk Of Bias Assessment Tool (PROBAST).Results Nineteen studies (6,851 patients) were included. Deep learning (DL) (nnU-Net, CNNs) identified PTE infiltration with high performance, achieving >80% sensitivity for recurrence (OR = 9.97), and outperforming radiomics (AUC = 0.61–0.98). Multimodal MRI (DTI/DSC+FLAIR) showed higher accuracy (0.90–0.98) than standard sequences. Synthetic MRI enabled standardized IDH/1p/19q prediction (AUC = 0.95/0.86). Voxel-wise histopathological validation was extremely rare (1 study).Discussion AI-based peritumoral phenotyping shows strong non-invasive potential. However, prospective clinical trials are essential to validate these predictive maps before their clinical use in surgical or radiotherapy planning.
Automated segmentation of unstained live cells in bright-field microscopy remains challenging for resource-constrained laboratories due to limited annotated data, computational resources, low contrast, motion blur, and overlapping cellular structures. We present a resource-efficient CNN pipeline comparing six frozen encoder backbones within enhanced U-Net architectures incorporating attention mechanisms, instance-aware modules, composite loss functions, hard-example mining, and 14-model ensemble voting. Trained on 857 bright-field and phase-contrast images expanded to 17,997 samples through augmentation, the framework required 6.5 hours on a Google Colab T4 GPU. External validation on LIVECell achieved Dice and F1 scores of 0.89, demonstrating strong cross-modality generalisation. On challenging bright-field images, the framework outperformed CellPose-SAM and StarDist while requiring substantially lower computational resources than foundation-model-based approaches. The proposed pipeline provides a reproducible and cost-effective solution for robust live-cell segmentation under constrained computational environments. 1
Liver ultrasound (US) is fundamental for screening and diagnosing hepatic diseases. This study introduces a weakly supervised deep learning approach for the identification of background liver conditions and focal liver lesions (FLLs) in US sequences. A retrospective dataset of 1,555 abdominal and liver US exams with exam-level labels was collected. Liver-containing images were automatically identified and cropped to the liver region using a YOLO11m-based classifier and detector. Each frame was then processed through a Swin Transformer encoder. Features were subsequently combined using Multi-Layer Perceptron (MLP) fusion to perform multi-label classification of clinical findings. Our model achieved strong performance for both background liver conditions: Cirrhosis (AUROC 0.962; Sensitivity 0.915; Specificity 0.906) and Fatty Liver (AUROC 0.928; Sensitivity 0.855; Specificity 0.844) and for FLLs: Any FLLs (AUROC 0.925; Sensitivity 0.895; Specificity 0.825), Cyst (AUROC 0.921; Sensitivity 0.869; Specificity 0.865), Nodule or Mass (AUROC 0.899; Sensitivity 0.957; Specificity 0.672), Focal Fatty Infiltration (AUROC 0.841; Sensitivity 0.566; Specificity 0.990) and Focal Fatty Sparing (AUROC 0.966; Sensitivity 0.916; Specificity 0.933). Despite relying only on exam-level labels, our weakly supervised framework achieved strong performance across multiple clinically relevant findings. Our pipeline effectively identifies FLLs from US sequences, highlighting its potential for future clinical validation and deployment as well as retrospective analysis of archived US exams.
This study proposes a hybrid teaching model that integrates a Fuzzy Neural Network (FNN) to improve body movement recognition and optimize exercise intensity in sports education. The model combines fuzzy limb-set representation with a multi-layer neural network to achieve accurate human action recognition, while a physical distance index is incorporated to enhance training sample selection. Experimental findings demonstrate a recognition accuracy of approximately 90.5%, with a maximum false recognition rate of 0.03 (one incorrect recognition in every 30 actions) across eight track-and-field activities. Furthermore, the proposed approach dynamically adjusts exercise intensity according to learners’ performance, resulting in a substantial improvement in students’ health status, increasing from 0.65 to 0.99. These findings suggest that adaptive, hybrid training strategies are more effective than fixed-intensity exercise programs in promoting physical fitness and learning outcomes. The results provide strong support for the application of intelligent teaching models in modern sports education.
We developed an interpretable machine learning model to classify cerebral small vessel disease (CSVD) using carotid plaque ultrasound radiomics combined with clinical features. In 193 participants (117 with CSVD, 76 controls), 107 ultrasound features were extracted, and 7 key features were selected via LASSO regression. Multimodal models significantly outperformed clinical-only models (p < 0.05), with the best model achieving an AUC of 0.888. SHAP analysis provided model interpretability, and ROC/decision curve analysis confirmed its robustness. This model offers a promising cost-effective CSVD screening tool for low-resource settings where MRI is limited. Further multicenter validation is recommended to enhance generalizability.
BCI clinical viability is hindered by "BCI illiteracy" (15-30% failure rate) and high latency. We propose CSOANet2026, a lightweight CNN with a FIR/ICA preprocessing pipeline, evaluated on 46 subjects. The 3,028-parameter model: (1) reaches 98.67% +/- 2.46% within-subject accuracy and 96.73% +/- 6.31% leave-one-subject-out (LOSO) accuracy, recovering decoding for "noisy" subjects who fail under linear baselines; (2) runs in 3.99 +/- 0.60 ms (CPU)/0.81 +/- 0.16 ms (GPU) per trial with a similar to 17 KB memory footprint; and (3) uses Grad-CAM and gate-weight analyses to verify reliance on neurophysiologically valid occipital Alpha-band (10-13 Hz) modulation-suppression during perception and synchronization during imagery. CSOANet2026 occupies a Pareto-optimal point of the accuracy-capacity frontier, matching deeper CNN baselines at 1/40 to 1/105 of their parameter count, while meeting real-time constraints for closed-loop neurofeedback.
Diabetic retinopathy (DR) is a leading cause of preventable blindness, requiring accurate severity grading for timely clinical intervention. Current automated grading methods face two key challenges: inadequate capture of regional lesion correlations, and poor performance on underrepresented severe grades due to class imbalance. This study developed and validated MRAMnet, a novel deep learning framework addressing these limitations through multi-region attention mechanisms and contrastive disentangled learning. Using two public fundus image datasets - APTOS2019 (3,662 images) and EYEPACS (35,126 images) - MRAMnet partitions fundus images into quadrants, processed through a dual-branch architecture combining channel-spatial attention and transformer mechanisms. A Decoupled Supervised Contrastive Loss (DSCL) function balances feature learning across all severity grades. Evaluated via 5-fold cross-validation, MRAMnet achieved classification accuracy of 89.07% (APTOS2019) and 90.31% (EYEPACS), representing a 3-5% improvement over state-of-the-art methods. For severe grades, F1-score improvements of 8.7% (severe non-proliferative DR) and 6.5% (proliferative DR) were observed. MRAMnet demonstrates strong potential clinical utility in supporting DR screening programmes, particularly for sight-threatening cases requiring urgent referral.
Early and accurate detection of leukaemia is essential for improving treatment outcomes; however, achieving high diagnostic reliability from high-dimensional microarray gene data remains challenging. This study investigates the application of optimised machine learning frameworks for leukaemia classification through systematic feature selection and classifier optimisation. Metaheuristic algorithms-whale optimisation algorithm, Grey Wolf Optimisation, Particle Swarm Optimisation and Flower Pollination Algorithm -were employed for gene subset extraction, followed by feature selection using Invasive Weed Optimisationand Harmonic Search. Multiple classifiers, including K-Nearest Neighbors, Decision Tree, Na & iuml;ve Bayes, Random Forest, and Support Vector Machine with linear, polynomial and RBF kernels, were evaluated using Accuracy, F1 score, MCC, Error Rate, Critical Success Index (CSI), G-mean and Cohen's Kappa. Among all the evaluated combinations, the PSO-HS-SVM (RBF) framework achieved superior performance. The findings demonstrate that the synergistic integration of optimisation-based feature selection with nonlinear SVM classifiers substantially enhances classification performance. These results highlight the potential of optimised machine learning approaches for reliable leukaemia detection while emphasising that the conclusions are limited to computational validation.
Objective verification of pedicle screw placement on postoperative imaging is essential for patient safety and for evaluating the biomechanical integrity of spinal constructs. Inaccurate screw positioning may compromise stability and increase the risk of neurological injury. Deep learning provides opportunities to automate segmentation of implants and surrounding anatomy, enabling objective and reproducible assessment of implant-anatomy relationships. This study evaluated the feasibility of training readily available convolutional neural networks (CNNs) to segment vertebrae, the spinal canal, and pedicle screws in postoperative thoracolumbar CT images using a limited dataset and 2D slice-based segmentation. One hundred manually annotated axial postoperative CT slices from 20 patients who had undergone posterior fixation were included and divided into training, validation, and test sets across two data splits. Four CNN architectures: YOLOv8, DeepLabv3+, U-net, and Attention U-net, were trained for multi-class segmentation. Model performance was assessed using class-wise Dice Similarity Coefficients (DSC). Across both splits, segmentation accuracy varied by model and anatomical class, mean DSC 0.62-0.90, reflecting variations across classes and models. DeepLabv3+achieved the highest overall performance, mean DSC 0.83 and 0.78. YOLOv8 demonstrated stable results across splits, particularly for screw segmentation. U-net and Attention U-net performed comparably for spinal canal and screw-head segmentation in the first split but showed reduced performance in the second split. Qualitative assessment confirmed that all models were able to identify key anatomical structures and implants, although metal artifacts affected performance. As a feasibility demonstration, the study showed that DeepLabv3+and YOLOv8 provided the most consistent results, indicating that robust automated screw-canal assessment is achievable even with limited clinical data.
Dysarthria in Parkinson's disease (PD) is a crucial early biomarker. While advanced deep learning models achieve high diagnostic accuracy, their large size and 'black-box' nature hinder clinical adoption on portable devices. This study proposes a novel interpretable lightweight framework for PD diagnosis via multi-task knowledge distillation. We first train a high capacity 'teacher' model (a deep FCNN) and transfer its knowledge to compact 'student' models using a composite loss function that distills output logits, feature representations, and attention. Experiments on a public dataset of 674 voice samples with 41 acoustic features rigorously compare our distilled models against nine traditional ML models and modern deep architectures. Results show: 1) The teacher model attains 90.6% accuracy, competitive with top traditional classifiers (e.g. KNN: 91.8%) and a state of the art deep model (93.0%). 2) The 3 layer student model achieves 13.2x parameter compression while preserving 97% of teacher accuracy (88.7% vs. 90.6%). 3) The 2 layer student reaches 70x compression, ideal for edge deployment. 4) Noise robustness tests confirm graceful performance degradation. 5) SHAP interpretability analysis shows the student retains the teacher's feature importance ranking (Kendall's tau = 0.82), validating clinical reasoning transfer. This work provides a practical framework for building efficient, accurate, and trustworthy AI tools for early PD screening.
Accurate measurement of lower-limb kinematics in outdoor sports is challenging due to the limitations of marker-based motion capture and inertial sensors. This study evaluates a stereo camera-based 3D pose estimation pipeline using real-time multi-person one-stage (RTMO) detection to reconstruct sagittal knee and hip angles during running, sprinting and jumping. Twelve adult football players were recorded with two synchronised GoPro cameras, while inertial measurement units (IMUs) provided reference joint angles. RTMO extracted 2D keypoints from both views, which were triangulated to 3D joint positions for angle computation. During running at 2-4 m, agreement with IMUs was strong (r > 0.90, RMSE < 10 degrees). At 8 m, accuracy declined (RMSE 23.74 degrees, r = 0.80). Sprinting produced higher errors than running and jumping yielded the lowest error (RMSE 5.11 degrees, r = 0.98). Findings indicate that stereo-based pose estimation can capture joint kinematics outdoors, with accuracy dependent on distance and activity type.
Femoral anteversion is a key morphological parameter of the human hip joint and varies widely across individuals. It is often considered mechanically destabilizing, based mainly on analyses of upright stance and walking. However, the hip is a three-dimensional load-bearing joint, and joint reaction force direction systematically changes with hip flexion. How anteversion interacts with this flexion-dependent load direction remains unclear. In this study, we develop a minimal three-dimensional geometric model to isolate this interaction. The femoral neck is modeled as a load-bearing axis defined by anteversion and neck-shaft (CCD) angle, while joint reaction force rotates posteriorly with increasing flexion. Alignment is quantified as the absolute dot product between neck axis and load direction. Results show that anteversion has no effect at neutral posture, decreases alignment in low-to-mid flexion, and partially restores alignment during deep flexion. These findings indicate that the mechanical role of femoral anteversion is inherently posture dependent.
We developed a deep learning (DL) algorithm to segment scaphoids and detect scaphoid fractures on single-incidence wrist radiographs, the most frequent carpal injuries whose early diagnosis is crucial for wrist function. We exploited a dataset of 1477 wrist radiographs and investigated strategies to mitigate data imbalance caused by low fracture prevalence (9%). The segmentation model achieved excellent precision (mAP@0.75 of 0.98), localising the scaphoid in all but one of 1141 test radiographs. However, severe class imbalance posed challenges in fracture detection. Our best fracture detection model achieved 74% sensitivity and 76% specificity on a balanced test dataset, surpassing an expert musculoskeletal radiologist’s sensitivity of 48% but falling short of their 94% specificity. This study demonstrates the potential of DL to detect scaphoid fractures from single-incidence radiographs, even in highly imbalanced datasets, potentially avoiding misdiagnosis and improving accuracy in emergency settings and non-specialised centres while reducing reliance on additional ionising imaging.
Ophthalmic diseases establish a vital global health concern, affecting millions of persons worldwide and affect vision loss which increases impact on the quality of life and economic burden. Classifying diseases is a crucial aspect of the ophthalmology, as it helps in diagnosis, treatment and research. It encompasses a wide range of conditions affecting the eyes and its surrounding. The classification can be based on anatomy, clinical presentation, age-related factors, systemic associations and traumatic causes. Ophthalmologists use these classifications to provide targeted treatments and interventions. In this review paper, we converse the impact of diseases on individuals and society, and image modalities used to classify and detect disease with severity. Additionally, this study explores risk factors along with the need for dedicated research, intervention strategies, diagnosis challenges and proposed system. By understanding the various circumstances, this study helps to diagnose the eye disease with its severity. This study reveals that 28% of researchers used the Messidor dataset with 24% applied CNN techniques, along with 57% diseases focusing on diabetic retinopathy. Despite existing methods, there is scope for improvement in detecting multiple ophthalmic diseases, which the proposed system aims to address by enhancing detection and classification with severity within a single framework.
Identification of a brain tumour is a very crucial task in medical diagnostics. It requires timely and accurate detection for effective treatment that improves the outcomes of patients. This challenge involves handling the variability and complexity of tumour imagery with high precision. The present work addresses these challenges by presenting a new framework that integrates TL with the Sugeno-Fuzzy Integral (SFI) for enhanced BT classification. Three different advanced TL models, namely ResNet-164, SqueezeNet, and DenseNet-201, will be used, combined with the SFI to aggregate their predictions for improving the accuracy of BT classification. For experimentation, two datasets are used: Dataset 1 consists of three tumour classes, namely glioma, pituitary, and meningioma, while Dataset 2 consists of four classes, adding the normal class. Preparation of the dataset included resizing, filtering, and balancing the dataset, followed by training and testing with a split of 70:30. Then, the developed model was tested on two MRI benchmark datasets, where classification accuracy of 99.19% was achieved for a three-class dataset and 98.92% for a four-class dataset, outperforming both individual and traditional ensemble models.
Accurate segmentation of the cytoplasm and nucleus in Pap smear images remains a challenging task for automated cervical cancer screening due to weak cell boundaries, overlapping structures, and staining variability. In this work, DeepSeg-Net, a novel deep learning-based segmentation framework, has been designed to robustly delineate cellular structures in cervical cytology images. The core novelty of DeepSeg-Net lies in its unified segmentation strategy that enhances feature representation while selectively focusing on diagnostically relevant regions, enabling reliable separation of cytoplasm and nucleus even under complex imaging conditions. By effectively suppressing background interference and strengthening multi-scale feature learning, the proposed approach addresses key limitations observed in existing segmentation methods. The performance of DeepSeg-Net is systematically evaluated against widely used U-Net-based architectures under identical experimental settings. The results demonstrate that the proposed method consistently outperforms conventional approaches in segmenting both cytoplasm and nucleus across cancerous and non-cancerous samples. Furthermore, the model exhibits strong generalisation capability when tested on an independent benchmark dataset with differing imaging characteristics. Overall, DeepSeg-Net offers a robust and effective segmentation backbone for Pap smear image analysis and has the potential to significantly support downstream cervical cancer screening and diagnostic workflows through accurate pixel-level cellular delineation.
A colorectal polyp is a small lesion that appears on the mucosal surface of the colon. Although many polyps are harmless, some can progress into colorectal cancer, becoming dangerous and potentially fatal when detected at advanced stages. Most existing research on polyp detection relies on deep learning methods, which, despite their effectiveness, suffer from several limitations, including high energy consumption, costly hardware requirements, and concerns related to privacy, security, and ethics. The proposed methodology introduces a low-cost alternative by separating each two-dimensional (2D) endoscopic image into its red, green, and blue (RGB) channels and converting each channel into a three-dimensional (3D) representation. A specialized colormap is applied to the 3D images, after which the colormap outputs of the green and blue layers are converted into binary images. Subtracting the colormap-based red layer from this combined binary image produces a mask that effectively isolates the polyp region. The method achieved strong results across three benchmark datasets. Recall reached 97% on CVC-ClinicDB, 95.5% on Kvasir-SEG, and 95.3% on CVC-ColonDB. Precision values were 96.5%, 93.4%, and 93.6%, respectively. F1-scores ranged from 94.5% to 95.3%. This approach offers an efficient, accessible solution suitable for devices with limited computational resources.
A significant rise in brain cancer cases over last few decades, insists early tumor detection, which is typically accomplished using vision-based computing and magnetic resonance imaging (MRI). Despite advances in deep learning for medical imaging, achieving accurate and timely tumor segmentation remains challenging due to issues like class imbalance, gradient vanishing, and premature convergence. Furthermore, limited works focus on simultaneous segmentation of the whole tumor and its substructures, which is essential for comprehensive diagnosis and treatment planning. To overcome these gaps, this paper proposes a Trans-Semantic Residual Deep Structure for Scalable Brain Tumor Segmentation namely BResUNet which is a combination of U-Net and ResUNet. This multi-stage deep network included sub-components like Pre-trained Activation Residual Layers (PARUL), attention gates, deep supervision, etc. The annotations followed by U-Net segmentation enable alleviating class imbalance, while Z-score normalization avoids the over-fitting problem. Architecturally, BResUNet is designed to perform segmentation of the whole tumor region as well as allied sub-structures as a multi-task problem. The ability to learn semantic contextual features as well as morphological features enable BResUNet to achieve optimal brain tumor segmentation results. To enable a scalable and time-efficient solution, BResUNet is executed over TensorFlow Distributed Computing Architecture with Multi-Task Mirroring (TD-CAM). Simulation results reveals the BResUNet’s higher performance with 95.37% accuracy, and 92.29% sensitivity, ensuring its applicability in real-time brain tumor diagnosis.
Modified abstractBone fractures of the hand are frequently missed in radiographic examinations due to their complex anatomy and subtle fracture patterns. This study proposes a hybrid deep learning-machine learning (DL-ML) framework for automated hand fracture detection using X-ray images. Pre-trained convolutional neural networks-ResNet-18, SqueezeNet, and AlexNet-were employed as feature extractors, while six machine learning classifiers, including Support Vector Machines, K-Nearest Neighbour, and Discriminant Analysis, were used for classification. Bayesian optimisation was applied to tune hyperparameters and enhance performance. The best results were achieved using ResNet-18 features combined with a KNN classifier, yielding an accuracy of 99.86%, precision of 0.997, recall of 0.997, and an F1-score of 0.997. In comparison, end-to-end deep learning models performed significantly worse, with AlexNet achieving a maximum accuracy of 78.06%. The results demonstrate that hybrid DL-ML approaches outperform standalone deep learning models and offer a lightweight, accurate solution to assist radiologists in reliable hand fracture diagnosis.
Alzheimer’s disease is a neurodegenerative brain disease that causes short-term memory loss. Recently, artificial intelligence (AI) is becoming important in the medical field and makes it possible to detect Alzheimer’s disease with greater accuracy. In this work, a novel Mamdani fuzzy inference system based on multi-textural biomarkers (MFIS-MTB) is proposed for the detection of Alzheimer stages. Initially, the MRI images are pre-processed to improve the quality of the image. The amygdala and hippocampus in the left hemisphere of the brain are segmented as the region of interest (ROI) from MRI scans of a small cohort. After normalising the ROI using µ±3σ, first-order histogram features (skewness and kurtosis) are retrieved. Additionally, two-dimensional wavelet features (max norm of the original image and diagonal detail coefficient) are extracted from the ROI. Based on these extracted textural markers, Mamdani fuzzy inference system (MFIS) is constructed with a least number of rules for the prognosis of the various phases of Alzheimer such as mild cognitive impairment (MCI), Alzheimer’s disease (AD) and normal controls (NC). The proposed MFIS-MTB model attains the accuracy of 96.13%, 94.73% and 93.11% for classifying AD vs NC, MCI vs NC and AD vs MCI respectively.
