Enhancing Pneumonia Detection in Pediatric Chest X-Rays Using CGAN-Augmented Datasets and Lightweight Deep Transfer Learning Models DOI Open Access
Coulibaly Mohamed,

Ronald Waweru Mwangi,

John Kihoro

et al.

Journal of Data Analysis and Information Processing, Journal Year: 2024, Volume and Issue: 12(01), P. 1 - 23

Published: Jan. 1, 2024

Pneumonia ranks as a leading cause of mortality, particularly in children aged five and under. Detecting this disease typically requires radiologists to examine chest X-rays report their findings physicians, task susceptible human error. The application Deep Transfer Learning (DTL) for the identification pneumonia through is hindered by shortage available images, which has led less than optimal DTL performance issues with overfitting. Overfitting characterized model’s learning that too closely fitted training data, reducing its effectiveness on unseen data. problem overfitting especially prevalent medical image processing due high costs extensive time required annotation, well challenge collecting substantial datasets also respect patient privacy concerning infectious diseases such pneumonia. To mitigate these challenges, paper introduces use conditional generative adversarial networks (CGAN) enrich dataset 2690 synthesized X-ray images minority class, aiming even out distribution improved diagnostic performance. Subsequently, we applied four modified lightweight deep transfer models Xception, MobileNetV2, MobileNet, EfficientNetB0. These have been fine-tuned evaluated, demonstrating remarkable detection accuracies 99.26%, 98.23%, 97.06%, 94.55%, respectively, across fifty epochs. experimental results validate proposed achieve accuracy rates, best model reaching up 99.26% effectiveness, outperforming other diagnosis from images.

Language: Английский

Integrated Model (IM- LTS) for Lung Tumor Segmentation using Neural Networks and IoMT]. DOI Creative Commons

J. Jayapradha,

Su-Cheng Haw, Palanichamy Naveen

et al.

MethodsX, Journal Year: 2025, Volume and Issue: 14, P. 103201 - 103201

Published: Feb. 7, 2025

Language: Английский

Citations

0
A Fully Automated, Expert-Perceptive Image Quality Assessment System for Whole-Body [18F]FDG PET/CT DOI Creative Commons
Cong Zhang, Xin Gao,

Xuebin Zheng

et al.

Research Square (Research Square), Journal Year: 2025, Volume and Issue: unknown

Published: March 26, 2025

Abstract Background The quality of clinical PET/CT images is critical for both accurate diagnosis and image-based research. However, current image assessment (IQA) methods predominantly rely on handcrafted features region-specific analyses, thereby limiting automation in whole-body multi-center evaluations. This study aims to develop an expert-perceptive deep learning-based IQA system [18F]FDG tackle the lack automated, interpretable assessments quality. Methods retrospective multicenter included scans from 718 patients. Automated identification localization algorithms were applied select predefined pairs PET CT slices images. Fifteen experienced experts, trained conduct blinded slice-level subjective assessments, provided average visual scores as reference standards. Using MANIQA framework, developed model integrates Vision Transformer, Transposed Attention, Scale Swin Transformer Blocks categorize into five classes. model’s correlation, consistency, accuracy with expert evaluations test sets statistically analysed assess system's performance. Additionally, model's ability distinguish high-quality was evaluated using receiver operating characteristic (ROC) curves. Results demonstrated high predicting categories showed strong concordance In across all body regions, achieved 0.832 0.902 CT. substantial agreement achieving Spearman coefficients (ρ) 0.891 0.624 CT, while Intraclass Correlation Coefficient (ICC) reached 0.953 0.92 discriminative performance, area under curve (AUC) ≥ 0.88 thoracic abdominal regions. Conclusions fully automated provides a robust comprehensive framework objective evaluation Furthermore, it demonstrates significant potential impartial, expert-level tool standardised IQA.

Language: Английский

Citations

0
Hybrid Classical–Quantum Transfer Learning for Cardiomegaly Detection in Chest X-rays DOI Creative Commons
Pierre Decoodt, Tan Jun Liang, Soham Bopardikar

et al.

Journal of Imaging, Journal Year: 2023, Volume and Issue: 9(7), P. 128 - 128

Published: June 25, 2023

Cardiovascular diseases are among the major health problems that likely to benefit from promising developments in quantum machine learning for medical imaging. The chest X-ray (CXR), a widely used modality, can reveal cardiomegaly, even when performed primarily non-cardiological indication. Based on pre-trained DenseNet-121, we designed hybrid classical–quantum (CQ) transfer models detect cardiomegaly CXRs. Using Qiskit and PennyLane, integrated parameterized circuit into classic network implemented PyTorch. We mined CheXpert public repository create balanced dataset with 2436 posteroanterior CXRs different patients distributed between control. k-fold cross-validation, CQ were trained using state vector simulator. normalized global effective dimension allowed us compare trainability run Qiskit. For prediction, ROC AUC scores up 0.93 accuracies 0.87 achieved several models, rivaling classical–classical (CC) model as reference. A trustworthy Grad-CAM++ heatmap hot zone covering heart was visualized more often QC option than CC (94% vs. 61%, p < 0.001), which may boost rate of acceptance by professionals.

Language: Английский

Citations

9
Integration of feature enhancement technique in Google inception network for breast cancer detection and classification DOI Creative Commons
Wasyihun Sema Admass, Yirga Yayeh Munaye, Ayodeji Olalekan Salau

et al.

Journal Of Big Data, Journal Year: 2024, Volume and Issue: 11(1)

Published: May 28, 2024

Abstract Breast cancer is a major public health concern, and early detection classification are essential for improving patient outcomes. However, breast tumors can be difficult to distinguish from benign tumors, leading high false positive rates in screening. The reason that both malignant have no consistent shape, found at the same position, variable sizes, correlations. ambiguity of correlation challenges computer-aided system, inconsistency morphology an expert identifying classifying what negative. Due this, most time, screen prone rates. This research paper presents introduction feature enhancement method into Google inception network classification. proposed model preserves local global information, which important addressing variability tumor their complex A locally preserving projection transformation function introduced retain information might lost intermediate output model. Additionally, transfer learning used improve performance on limited datasets. evaluated dataset ultrasound images achieves accuracy 99.81%, recall 96.48%, sensitivity 93.0%. These results demonstrate effectiveness

Language: Английский

Citations

3
Explainability of deep learning models in medical video analysis: a survey DOI Creative Commons
Michal Kolárik, Martin Sarnovský, Ján Paralič

et al.

PeerJ Computer Science, Journal Year: 2023, Volume and Issue: 9, P. e1253 - e1253

Published: March 14, 2023

Deep learning methods have proven to be effective for multiple diagnostic tasks in medicine and been performing significantly better comparison other traditional machine methods. However, the black-box nature of deep neural networks has restricted their use real-world applications, especially healthcare. Therefore, explainability models, which focuses on providing comprehensible explanations model outputs, may affect possibility adoption such models clinical use. There are various studies reviewing approaches domains. This article provides a review current applications explainable specific area medical data analysis-medical video processing tasks. The introduces field AI summarizes most important requirements applications. Subsequently, we provide an overview existing methods, evaluation metrics focus more those that can applied analytical involving domain. Finally identify some open research issues analysed area.

Language: Английский

Citations

7
Recurrence quantification analysis of rs-fMRI data: A method to detect subtle changes in the TgF344-AD rat model DOI
Arash Rezaei -, Monica van den Berg, Hajar Mirlohi

et al.

Computer Methods and Programs in Biomedicine, Journal Year: 2024, Volume and Issue: 257, P. 108378 - 108378

Published: Aug. 16, 2024

Language: Английский

Citations

2
LSTM Multi-Stage Transfer Learning for Blood Pressure Estimation Using Photoplethysmography DOI Open Access
Noor Faris Ali, Mohamed Atef

Electronics, Journal Year: 2022, Volume and Issue: 11(22), P. 3749 - 3749

Published: Nov. 15, 2022

Considerable research has been devoted to developing machine-learning models for continuous Blood Pressure (BP) estimation. A challenging problem that arises in this domain is the selection of optimal features with interpretable medical professionals. The aim study was investigate evidence-based physiologically motivating based on a solid physiological background BP determinants. powerful and compact set encompassing six oriented extracted addition another consisting commonly used comparison purposes. In study, we proposed predictive model using Long Short-Term Memory (LSTM) networks multi-stage transfer learning approach. topology consists three cascaded stages. First, classification stage. Second, Mean Arterial (MAP) regression stage further approximate quantity proportional Vascular Resistance (VR) Cardiac Output (CO) from PPG signal. Third, main estimation final (final prediction) able exploit embedded correlations between along derived outputs carrying hemodynamic characteristics through sub-sequence We also constructed traditional single-stage Artificial Neural Network (ANN) LSTM-based appraise performance gain our model. were tested evaluated 40 subjects MIMIC II database. attained MAE ± SD 2.03 3.12 SBP 1.18 1.70 mmHg DBP. resulted drastic error reduction, up 86.21%, compared trained features. superior provides confirmatory evidence selected transferable among stages coupled high-performing enhance blood pressure accuracy signals. This indicates compelling nature sufficiency efficient set.

Language: Английский

Citations

11
Increasing Robustness of Intracortical Brain-Computer Interfaces for Recording Condition Changes via Data Augmentation DOI
Shih‐Hung Yang,

Chun-Jui Huang,

Jhih-Siang Huang

et al.

Computer Methods and Programs in Biomedicine, Journal Year: 2024, Volume and Issue: 251, P. 108208 - 108208

Published: May 3, 2024

Language: Английский

Citations

1
Transfer Learning Video Classification of Preserved, Mid-Range, and Reduced Left Ventricular Ejection Fraction in Echocardiography DOI Creative Commons
Pierre Decoodt, Daniel Sierra-Sosa, Laura Anghel

et al.

Diagnostics, Journal Year: 2024, Volume and Issue: 14(13), P. 1439 - 1439

Published: July 5, 2024

Identifying patients with left ventricular ejection fraction (EF), either reduced [EF < 40% (rEF)], mid-range 40-50% (mEF)], or preserved > 50% (pEF)], is considered of primary clinical importance. An end-to-end video classification using AutoML in Google Vertex AI was applied to echocardiographic recordings. Datasets balanced by majority undersampling, each corresponding one out three possible classifications, were obtained from the Standford EchoNet-Dynamic repository. A train-test split 75/25 applied. binary rEF vs. not demonstrated good performance (test dataset: ROC AUC score 0.939, accuracy 0.863, sensitivity 0.894, specificity 0.831, positive predicting value 0.842). second pEF slightly less performing 0.917, 0.829, 0.761, 0.891, 0.888). ternary also explored, and lower observed, mainly for mEF class. non-AutoML PyTorch implementation open access confirmed feasibility our approach. With this proof concept, based on transfer learning categorize EF merits consideration further evaluation prospective studies.

Language: Английский

Citations

1
A Combined MobileNetV2 and CBAM Model to Improve Classifying the Breast Cancer Ultrasound Images DOI Creative Commons

Muhammad Rakha,

Mahmud Dwi Sulistiyo, Dewi Nasien

et al.

Journal of Applied Engineering and Technological Science (JAETS), Journal Year: 2024, Volume and Issue: 6(1), P. 561 - 578

Published: Dec. 15, 2024

Breast cancer is the main cause of death in women throughout world. Early detection using ultrasound very necessary to reduce cases breast cancer. However, analysis process requires a lot time and medical personnel because classification difficult due noise, complex texture, subjective assessment. Previous studies were successful but required large computations models. This research aims overcome these shortcomings by lighter more accurate model. We integrated CBAM attention module into MobileNetV2 model improve accuracy, speed up diagnosis, computational requirements. Gradient Weighted Class Activation Mapping (Grad-CAM) used explanations. Ultrasound images from two databases combined train, validate, test this The results show that MobileNetV2-CBAM achieves accuracy 93%, higher than models VGG-16 (80%), VGG-19 (82%), InceptionV3 ResNet-50 (84%). proven performance with an 11% increase accuracy. Grad-CAM visualization shows can better focus on localizing important regions images, providing clearer explanations assisting diagnosis.

Language: Английский

Citations

1