Medical image segmentation based on frequency domain decomposition SVD linear attention

Scientific Reports, Год журнала: 2025, Номер 15(1)

Опубликована: Янв. 22, 2025

Convolutional Neural Networks (CNNs) have achieved remarkable segmentation accuracy in medical image tasks. However, the Vision Transformer (ViT) model, with its capability of extracting global information, offers a significant advantage contextual information compared to limited receptive field convolutional kernels CNNs. Despite this, ViT models struggle fully detect and extract high-frequency signals, such as textures boundaries, images. These features are essential imaging, targets like tumors pathological organs exhibit differences texture boundaries across different stages. Additionally, high resolution images leads computational complexity self-attention mechanism ViTs. To address these limitations, we propose network framework based on frequency domain decomposition using Laplacian pyramid. This approach selectively computes attention for signals original enhance spatial structural effectively. During feature computation, introduce Singular Value Decomposition (SVD) an effective representation matrix from image, which is then applied computation process linear projection. method reduces while preserving features. We demonstrated validity superiority our model Abdominal Multi-Organ Segmentation dataset Dermatological Disease dataset, Synapse score 82.68 Dice metrics 17.23 mm HD metrics. Experimental results indicate that consistently exhibits effectiveness improved various datasets.

Язык: Английский

---

CerviLearnNet: Advancing cervical cancer diagnosis with reinforcement learning-enhanced convolutional networks

Heliyon, Год журнала: 2024, Номер 10(9), С. e29913 - e29913

Опубликована: Апрель 24, 2024

Women tend to face many problems throughout their lives; cervical cancer is one of the most dangerous diseases that they can face, and it has negative consequences. Regular screening treatment precancerous lesions play a vital role in fight against cancer. It becoming increasingly common medical practice predict early stages serious illnesses, such as heart attacks, kidney failure, cancer, using machine learning-based techniques. To overcome these obstacles, we propose use auxiliary modules special residual block, record contextual interactions between object classes support reference strategy. Unlike latest state-of-the-art classification method, create new architecture called Reinforcement Learning Cancer Network, "RL-CancerNet", which diagnoses with incredible accuracy. We trained tested our method on two well-known publicly available datasets, SipaKMeD Herlev, assess enable comparisons earlier methods. Cervical images were labeled this dataset; therefore, had be marked manually. Our study shows that, compared previous approaches for assignment classifying an cellular change, proposed approach generates more reliable stable image derived from datasets vastly different sizes, indicating will effective other datasets.

Язык: Английский

---

Simplified Knowledge Distillation for Deep Neural Networks Bridging the Performance Gap with a Novel Teacher–Student Architecture

Electronics, Год журнала: 2024, Номер 13(22), С. 4530 - 4530

Опубликована: Ноя. 18, 2024

The rapid evolution of deep learning has led to significant achievements in computer vision, primarily driven by complex convolutional neural networks (CNNs). However, the increasing depth and parameter count these often result overfitting elevated computational demands. Knowledge distillation (KD) emerged as a promising technique address issues transferring knowledge from large, well-trained teacher model more compact student model. This paper introduces novel method that simplifies process narrows performance gap between models without relying on intricate representations. Our approach leverages unique network architecture designed enhance efficiency effectiveness transfer. Additionally, we introduce streamlined transfers effectively through simplified process, enabling achieve high accuracy with reduced Comprehensive experiments conducted CIFAR-10 dataset demonstrate our proposed achieves superior compared traditional KD methods established architectures such ResNet VGG networks. not only maintains but also significantly reduces training validation losses. Key findings highlight optimal hyperparameter settings (temperature T = 15.0 smoothing factor α 0.7), which yield highest lowest loss values. research contributes theoretical practical advancements distillation, providing robust framework for future applications compression optimization. simplicity pave way accessible scalable solutions deployment.

Язык: Английский

---

Efficient image classification through collaborative knowledge distillation: A novel AlexNet modification approach

Heliyon, Год журнала: 2024, Номер 10(14), С. e34376 - e34376

Опубликована: Июль 1, 2024

This paper introduces an innovative image classification technique utilizing knowledge distillation, tailored for a lightweight model structure. The core of the approach is modified version AlexNet architecture, enhanced with depthwise-separable convolution layers. A unique aspect this work Teacher-Student Collaborative Knowledge Distillation (TSKD) method. Unlike conventional distillation techniques, TSKD employs dual-layered learning strategy, where student learns from both final output and intermediate layers teacher model. collaborative enables to actively engage in process, resulting more efficient transfer. emphasizes suitability scenarios limited computational resources. achieved through architectural optimizations introduction specialized loss functions, which balance trade-off between complexity efficiency. study demonstrates that despite its nature, maintains high accuracy robustness tasks. Key contributions include use AlexNet, transfer, development functions. These advancements collectively contribute effectiveness environments constraints, making it valuable contribution field classification.

Язык: Английский

---

Lightweight YOLOv8 for tongue teeth marks and fissures detection based on C2f_DCNv3

Scientific Reports, Год журнала: 2025, Номер 15(1)

Опубликована: Янв. 10, 2025

This paper propose a significantly enhanced YOLOv8 model specifically designed for detecting tongue fissures and teeth marks in Traditional Chinese Medicine (TCM) diagnostic images. By integrating the C2f_DCNv3 module, which incorporates Deformable Convolutions (DCN), replace original C2f enabling to exhibit exceptional adaptability intricate irregular features, such as fine marks. Furthermore, introduction of Squeeze-and-Excitation (SE) attention mechanism optimizes feature weighting, allowing focus more accurately on key regions image, even presence complex backgrounds. The proposed demonstrates significant performance improvement, achieving an average precision (mAP) 92.77%, substantial enhancement over YOLOv8. Additionally, reduces computational cost by approximately one-third terms FLOPS, maintaining high accuracy while greatly decreasing number parameters, thus offering robust resource-efficient solution. For crack detection, mAP increases 91.34%, with notable improvements F1 score, precision, recall. Teeth mark detection also sees boost, 94.21%. These advancements underscore model's outstanding TCM image analysis, providing accurate, efficient, reliable tool clinical applications.

Язык: Английский

---

The Diagnostic Classification of the Pathological Image Using Computer Vision

Algorithms, Год журнала: 2025, Номер 18(2), С. 96 - 96

Опубликована: Фев. 8, 2025

Computer vision and artificial intelligence have revolutionized the field of pathological image analysis, enabling faster more accurate diagnostic classification. Deep learning architectures like convolutional neural networks (CNNs), shown superior performance in tasks such as classification, segmentation, object detection pathology. has significantly improved accuracy disease diagnosis healthcare. By leveraging advanced algorithms machine techniques, computer systems can analyze medical images with high precision, often matching or even surpassing human expert performance. In pathology, deep models been trained on large datasets annotated pathology to perform cancer diagnosis, grading, prognostication. While approaches show great promise challenges remain, including issues related model interpretability, reliability, generalization across diverse patient populations imaging settings.

Язык: Английский

---

Joint Driver State Classification Approach: Face Classification Model Development and Facial Feature Analysis Improvement

Sensors, Год журнала: 2025, Номер 25(5), С. 1472 - 1472

Опубликована: Фев. 27, 2025

Driver drowsiness remains a critical factor in road safety, necessitating the development of robust detection methodologies. This study presents dual-framework approach that integrates convolutional neural network (CNN) and facial landmark analysis model to enhance detection. The CNN classifies driver states into “Awake” “Drowsy”, achieving classification accuracy 92.5%. In parallel, deep learning-based analyzes driver’s physiological state by extracting analyzing features. model’s was significantly enhanced through advanced image preprocessing techniques, including normalization, illumination correction, face hallucination, reaching 97.33% accuracy. proposed dual-model architecture leverages imagery detect key indicators, such as eye closure dynamics, yawning patterns, head movement trajectories. By integrating CNN-based with precise analysis, this not only improves robustness but also ensures greater resilience under challenging conditions, low-light environments. findings underscore efficacy multi-model approaches their potential for real-world implementation safety mitigate drowsiness-related vehicular accidents.

Язык: Английский

---

Vision Mamba and xLSTM-UNet for medical image segmentation

Scientific Reports, Год журнала: 2025, Номер 15(1)

Опубликована: Март 10, 2025

Deep learning-based medical image segmentation methods are generally divided into convolutional neural networks (CNNs) and Transformer-based models. Traditional CNNs limited by their receptive field, making it challenging to capture long-range dependencies. While Transformers excel at modeling global information, high computational complexity restricts practical application in clinical scenarios. To address these limitations, this study introduces VMAXL-UNet, a novel network that integrates Structured State Space Models (SSM) lightweight LSTMs (xLSTM). The incorporates Visual (VSS) ViL modules the encoder efficiently fuse local boundary details with semantic context. VSS module leverages SSM dependencies extract critical features from distant regions. Meanwhile, employs gating mechanism enhance integration of features, thereby improving accuracy robustness. Experiments on datasets such as ISIC17, ISIC18, CVC-ClinicDB, Kvasir demonstrate VMAXL-UNet significantly outperforms traditional models capturing lesion boundaries correlations. These results highlight model's superior performance provide promising approach for efficient complex imaging

Язык: Английский

---

Research on recognition of diabetic retinopathy hemorrhage lesions based on fine tuning of segment anything model

Scientific Reports, Год журнала: 2025, Номер 15(1)

Опубликована: Март 25, 2025

Diabetic Retinopathy (DR) demands precise hemorrhage detection for early diagnosis, yet manual identification faces challenges due to hemorrhagic lesions' varied sizes, complex shapes, and color similarities surrounding tissues, which obscure boundaries reduce contrast. To address this, we propose SAM-ada-Res, a novel dual-encoder model integrating pre-trained Segment Anything Model (SAM) ResNet101. SAM captures global semantic context distinguish ambiguous lesions from vessels, while ResNet101 extracts fine-grained details through its deep hierarchical layers. Feature maps both encoders are fused via channel-wise concatenation, enabling the decoder localize with high precision. A lightweight Adapter fine-tunes retinal tasks without retraining backbone, ensuring task-specific adaptation. Evaluated on three datasets (OIA-DDR, IDRiD, JYFY-HE), SAM-ada-Res outperforms state-of-the-art methods in nDice (0.6040 JYFY-HE) nIoU (0.4182 IDRiD), demonstrating superior generalization robustness. An online platform further streamlines clinical deployment, enhancing diagnostic efficiency. By synergizing SAM's generalizable vision capabilities ResNet's localized feature extraction, overcomes key DR detection, offering robust tool intervention. This work bridges technical innovation practicality, advancing automated diagnosis.

Язык: Английский

---

Machine Learning and Deep Learning Approaches in Lifespan Brain Age Prediction: A Comprehensive Review

Tomography, Год журнала: 2024, Номер 10(8), С. 1238 - 1262

Опубликована: Авг. 12, 2024

The concept of 'brain age', derived from neuroimaging data, serves as a crucial biomarker reflecting cognitive vitality and neurodegenerative trajectories. In the past decade, machine learning (ML) deep (DL) integration has transformed field, providing advanced models for brain age estimation. However, achieving precise prediction across all ages remains significant analytical challenge. This comprehensive review scrutinizes advancements in ML- DL-based prediction, analyzing 52 peer-reviewed studies 2020 to 2024. It assesses various model architectures, highlighting their effectiveness nuances lifespan studies. By comparing ML DL, strengths forecasting methodological limitations are revealed. Finally, key findings reviewed articles summarized number major issues related ML/DL-based discussed. Through this study, we aim at synthesis current state emphasizing both persistent challenges, guiding future research, technological advancements, improving early intervention strategies diseases.

Язык: Английский

---