Explainable machine learning for assessing upper respiratory tract of racehorses from endoscopy videos DOI Creative Commons
Anas Tahir,

Li Guo,

Rabab Ward

et al.

Computers in Biology and Medicine, Journal Year: 2024, Volume and Issue: 181, P. 109030 - 109030

Published: Aug. 22, 2024

Laryngeal hemiplegia (LH) is a major upper respiratory tract (URT) complication in racehorses. Endoscopy imaging of horse throat gold standard for URT assessment. However, current manual assessment faces several challenges, stemming from the poor quality endoscopy videos and subjectivity grading. To overcome such limitations, we propose an explainable machine learning (ML)-based solution efficient Specifically, cascaded YOLOv8 architecture utilized to segment key semantic regions landmarks per frame. Several spatiotemporal features are then extracted points fed decision tree (DT) model classify LH as Grade 1,2,3 or 4 denoting absence LH, mild, moderate, severe respectively. The proposed method, validated through 5-fold cross-validation on 107 videos, showed promising performance classifying different grades with 100%, 91.18%, 94.74% 100% sensitivity values 1 4, Further validation external dataset 72 cases confirmed its generalization capability 90%, 80.95%, We introduced explainability related functions, including: (i) visualization output detect landmark estimation errors which can affect final classification, (ii) time-series assess video quality, (iii) backtracking DT identify borderline cases. incorporated domain knowledge (e.g., veterinarian diagnostic procedures) into ML framework. This provides assistive tool clinical-relevance that ease speed up by veterinarians.

Language: Английский

Visionary vigilance: Optimized YOLOV8 for fallen person detection with large-scale benchmark dataset DOI
Habib Ullah Khan, Inam Ullah, Mohammad Shabaz

et al.

Image and Vision Computing, Journal Year: 2024, Volume and Issue: 149, P. 105195 - 105195

Published: July 27, 2024

Language: Английский

Citations

25

Real-Time Household Waste Detection and Classification for Sustainable Recycling: A Deep Learning Approach DOI Open Access

Ali Arishi

Sustainability, Journal Year: 2025, Volume and Issue: 17(5), P. 1902 - 1902

Published: Feb. 24, 2025

As global waste production continues to rise, improper handling of household significantly contributes environmental pollution and resource depletion. Inefficient sorting at the level leads contamination recyclables, reducing recycling efficiency increasing landfill waste. Effective is essential for conserving manual labor, protecting environment, ensuring sustainable development human progress. Recently, advancements in deep learning computer vision have offered a promising pathway improve process, though significant developmental steps are still required. Enhancing automated detection classification through could bring substantial societal benefits. However, classifying identifying materials presents challenges due complex diverse nature waste, coupled with limited availability data on management. This paper real-time system based YOLOv8 model, designed enhance processes level. The proposed detects classifies range items. Experiments were conducted custom dataset comprising 3775 images across 17 types common one-stage model demonstrated superior performance, outperforming traditional two-stage detectors. To accuracy robustness original YOLOv8, five augmentation techniques two attention mechanisms incorporated. Notably, enhanced YOLOv8-CBAM achieved mean average precision (mAP) 89.5%, improvement 4.2% increase over baseline model. methodology improvements applied provide more efficient effective AI framework applications smart bins, robotic pickers, large-scale systems.

Language: Английский

Citations

1

Object Detection for Brain Cancer Detection and Localization DOI Creative Commons
Francesco Mercaldo, Luca Brunese, Fabio Martinelli

et al.

Applied Sciences, Journal Year: 2023, Volume and Issue: 13(16), P. 9158 - 9158

Published: Aug. 11, 2023

Brain cancer is acknowledged as one of the most aggressive tumors, with a significant impact on patient survival rates. Unfortunately, approximately 70% patients diagnosed this malignant do not survive. This paper introduces method designed to detect and localize brain by proposing an automated approach for detection localization cancer. The utilizes magnetic resonance imaging analysis. By leveraging information provided medical images, proposed aims enhance precise improve prognosis treatment outcomes patients. We exploit YOLO model automatically cancer: in analysis 300 images we obtain precision 0.943 recall 0.923 while, relating localization, mAP_0.5 equal 0.941 reached, thus showing effectiveness localization.

Language: Английский

Citations

12

KONet: Toward a Weighted Ensemble Learning Model for Knee Osteoporosis Classification DOI Creative Commons
M. J. Aashik Rasool, Shabir Ahmad, Sabina Umirzakova

et al.

IEEE Access, Journal Year: 2024, Volume and Issue: 12, P. 5731 - 5742

Published: Jan. 1, 2024

Knee osteoporosis (KOP) is a skeletal disorder characterized by bone tissue degradation and low density, leading to high risk of fractures in the knee area. The traditional method for identifying radiography, which requires sufficient expertise from specialists. However, sheer volume X-rays subtle variations among them may lead misinterpretation. In recent years, deep learning algorithms have revolutionized medical diagnosis reduced misclassification. Specifically, convolutional neural network (CNN)-based been utilized automate diagnostic process as they inherent ability extract important features that are difficult identify manually. relying on single result suboptimal performance, ineffective deployment domain. To alleviate this issue, study, we propose robust detection method, KONet, utilizes weighted ensemble approach distinguish between normal osteoporotic conditions, even when there minor data. validate architectural choices approach, conducted experiments various state-of-the-art CNN-based models using transfer learning. Extensive indicated proposed model achieves higher accuracy than existing models, outperforming significant margin.

Language: Английский

Citations

4

Colorectal image analysis for polyp diagnosis DOI Creative Commons
Pengcheng Zhu, Jingjing Wan, Wei Shao

et al.

Frontiers in Computational Neuroscience, Journal Year: 2024, Volume and Issue: 18

Published: Feb. 9, 2024

Colorectal polyp is an important early manifestation of colorectal cancer, which significant for the prevention cancer. Despite timely detection and manual intervention polyps can reduce their chances becoming cancerous, most existing methods ignore uncertainties location problems polyps, causing a degradation in performance. To address these problems, this paper, we propose novel image analysis method diagnosis via PAM-Net. Specifically, parallel attention module designed to enhance images improving certainties polyps. In addition, our introduces GWD loss accuracy from perspective location. Extensive experimental results demonstrate effectiveness proposed compared with SOTA baselines. This study enhances performance contributes clinical medicine.

Language: Английский

Citations

4

Practical X-ray gastric cancer diagnostic support using refined stochastic data augmentation and hard boundary box training DOI

Hideaki Okamoto,

Quan Huu,

Takakiyo Nomura

et al.

Artificial Intelligence in Medicine, Journal Year: 2025, Volume and Issue: 161, P. 103075 - 103075

Published: Feb. 1, 2025

Language: Английский

Citations

0

TQGDNet: Coronary Artery Calcium Deposit Detection on Computed Tomography DOI Creative Commons
Wei-Chien Wang, Christopher Yu,

Euijoon Ahn

et al.

Computerized Medical Imaging and Graphics, Journal Year: 2025, Volume and Issue: 121, P. 102503 - 102503

Published: Feb. 6, 2025

Coronary artery disease (CAD) continues to be a leading global cause of cardiovascular related mortality. The scoring coronary calcium (CAC) using computer tomography (CT) images is diagnostic instrument for evaluating the risk asymptomatic individuals prone atherosclerotic disease. State-of-the-art automated CAC methods rely on large annotated datasets train convolutional neural network (CNN) models. However, these do not integrate features across different levels and layers CNN, particularly in lower where important information regarding small regions are present. In this study, we propose new CNN model specifically designed effectively capture associated with their surrounding areas low-contrast CT images. Our integrates detection module two fusion modules focusing connect more deeper wider neurons (or nodes) multiple adjacent levels. first module, called ThrConvs, includes three convolution blocks tailored detecting objects characterized by low contrast. Following this, introduced: (i) Queen-fusion (Qf), which introduces cross-scale feature method fuse from and, (ii) lower-layer Gather-and-Distribute (GD) focuses learning comprehensive small-sized deposits surroundings. We demonstrate superior performance our public OrCaScore dataset, encompassing 269 deposits, surpassing capabilities previous state-of-the-art works. enhanced approach, achieving notable 2.3-3.6 % improvement mean Pixel Accuracy (mPA) both private Concord dataset established methods.

Language: Английский

Citations

0

PixelCraftSR: Efficient Super-Resolution with Multi-Agent Reinforcement for Edge Devices DOI Creative Commons
M. J. Aashik Rasool,

Shabir Ahmed,

S. M. A. Sharif

et al.

Sensors, Journal Year: 2025, Volume and Issue: 25(7), P. 2242 - 2242

Published: April 2, 2025

Single-image super-resolution imaging methods are increasingly being employed owing to their immense applicability in numerous domains, such as medical imaging, display manufacturing, and digital zooming. Despite widespread usability, the existing learning-based (SR) computationally expensive inefficient for resource-constrained IoT devices. In this study, we propose a lightweight model based on multi-agent reinforcement-learning approach that employs multiple agents at pixel level construct images by following asynchronous actor-critic policy. The iteratively select predefined set of actions be executed within five time steps new image state, followed action maximizes cumulative reward. We thoroughly evaluate compare our proposed method with methods. Experimental results illustrate can outperform models both qualitative quantitative scores despite having significantly less computational complexity. practicability is confirmed further evaluating it platforms, including edge

Language: Английский

Citations

0

YOLO-SLD: An Attention Mechanism-Improved YOLO for License Plate Detection DOI Creative Commons
Ming‐An Chung,

Yu‐Jou Lin,

Chia-Wei Lin

et al.

IEEE Access, Journal Year: 2024, Volume and Issue: 12, P. 89035 - 89045

Published: Jan. 1, 2024

The vehicle license plate detection plays a key role in Intelligent Transportation Systems. Detecting plates, such as cars, trucks, and vans, is useful for law enforcement, surveillance, toll booth operations. How to detect plates quickly accurately crucial recognition. However, the uneven light condition or oblique shooting angle of be detected changes dramatically real-world complex capture scenarios difficulty increases. At same time, distance, lighting, angle, other requirements are quite high, which seriously affects performance. Therefore, an improved YOLOv7 integrating parameter-free attention module SimAM was proposed, namely YOLO-SLD. Without modifying original ELAN architecture, component YOLOv7, mechanism added at end better extract features increase computational efficiency. More importantly, does not require any parameters network, reducing model computation, simplifying calculation process. performance with different mechanisms tested on CCPD dataset first time proposed method proven effective. experimental result shows that YOLO-SLD has higher accuracy more lightweight mAP 0.5 overall improvement from 98.44% 98.91%, 0.47% accuracy. test subset dark images 93.5% 96.7%, 3.2% parameter size reduced by 1.2 million compared model. Its than prevalent algorithms.

Language: Английский

Citations

3

Using AI Segmentation Models to Improve Foreign Body Detection and Triage from Ultrasound Images DOI Creative Commons
Lawrence Holland, Sofia I. Hernandez Torres, Eric J. Snider

et al.

Bioengineering, Journal Year: 2024, Volume and Issue: 11(2), P. 128 - 128

Published: Jan. 29, 2024

Medical imaging can be a critical tool for triaging casualties in trauma situations. In remote or military medicine scenarios, triage is essential identifying how to use limited resources prioritize evacuation the most serious cases. Ultrasound imaging, while portable and often available near point of injury, only used if images are properly acquired, interpreted, objectively scored. Here, we detail AI segmentation models improving image interpretation objective evaluation medical application focused on foreign bodies embedded tissues at variable distances from neurovascular features. previously collected tissue phantom with without features were labeled ground truth masks. These sets train two different frameworks: YOLOv7 U-Net models. Overall, both approaches successful shrapnel set, outperforming single-class segmentation. Both also evaluated more complex set containing shrapnel, artery, vein, nerve obtained higher precision scores across multiple classes whereas achieved recall scores. Using each model, distance metric was adapted measure proximity nearest feature, closely mirroring measured labels. detecting ultrasound could allow improved injury emergency scenarios.

Language: Английский

Citations

2