Neural Architecture Search for biomedical image classification: A comparative study across data modalities

Artificial Intelligence in Medicine, Journal Year: 2025, Volume and Issue: 160, P. 103064 - 103064

Published: Jan. 5, 2025

Language: Английский

---

Continual Learning in Medicine: A Systematic Literature Review

Neural Processing Letters, Journal Year: 2025, Volume and Issue: 57(1)

Published: Jan. 7, 2025

Abstract Continual Learning (CL) is a novel AI paradigm in which tasks and data are made available over time; thus, the trained model computed on basis of stream data. CL-based approaches able to learn new skills knowledge without forgetting previous ones, with no guaranteed access previously encountered data, mitigating so-called “catastrophic forgetting” phenomenon. Interestingly, by making systems improve time need for large amounts or computational resources, CL can help at reducing impact computationally-expensive energy-intensive activities; hence, play key role path towards more green AIs, enabling efficient sustainable uses resources. In this work, we describe different methods proposed literature solve tasks; survey applications, highlighting strengths weaknesses, particular focus biomedical context. Furthermore, discuss how make robust suitable wider range applications.

Language: Английский

---

DeepLeaf: an optimized deep learning approach for automated recognition of grapevine leaf diseases

Neural Computing and Applications, Journal Year: 2025, Volume and Issue: unknown

Published: Feb. 28, 2025

Language: Английский

---

A novel lightweight CNN for chest X-ray-based lung disease identification on heterogeneous embedded system

Applied Intelligence, Journal Year: 2024, Volume and Issue: 54(6), P. 4756 - 4780

Published: March 1, 2024

Abstract The global spread of epidemic lung diseases, including COVID-19, underscores the need for efficient diagnostic methods. Addressing this, we developed and tested a computer-aided, lightweight Convolutional Neural Network (CNN) rapid accurate identification diseases from 29,131 aggregated Chest X-ray (CXR) images representing seven disease categories. Employing five-fold cross-validation method to ensure robustness our results, CNN model, optimized heterogeneous embedded devices, demonstrated superior performance. It achieved 98.56% accuracy, outperforming established networks like ResNet50, NASNetMobile, Xception, MobileNetV2, DenseNet121, ViT-B/16 across precision, recall, F1-score, AUC metrics. Notably, model requires significantly less computational power only 55 minutes average training time per fold, making it highly suitable resource-constrained environments. This study contributes developing efficient, in medical image analysis, underscoring their potential enhance point-of-care processes.

Language: Английский

---

A review of convolutional neural network based methods for medical image classification

Computers in Biology and Medicine, Journal Year: 2024, Volume and Issue: 185, P. 109507 - 109507

Published: Dec. 3, 2024

Language: Английский

---

OCTNet: A Modified Multi-Scale Attention Feature Fusion Network with InceptionV3 for Retinal OCT Image Classification

Mathematics, Journal Year: 2024, Volume and Issue: 12(19), P. 3003 - 3003

Published: Sept. 26, 2024

Classification and identification of eye diseases using Optical Coherence Tomography (OCT) has been a challenging task trending research area in recent years. Accurate classification detection different are crucial for effective care management improving vision outcomes. Current methods fall into two main categories: traditional deep learning-based approaches. Traditional approaches rely on machine learning feature extraction, while utilize data-driven model training. In years, Deep Learning (DL) Machine (ML) algorithms have become essential tools, particularly medical image classification, widely used to classify identify various diseases. However, due the high spatial similarities OCT images, accurate remains task. this paper, we introduce novel called “OCTNet” that integrates combining InceptionV3 with modified multi-scale attention-based attention block enhance performance. OCTNet employs an backbone fusion dual modules construct proposed architecture. The generates rich features from capturing both local global aspects, which then enhanced by utilizing block, resulting significantly improved map. To evaluate model’s performance, utilized state-of-the-art (SOTA) datasets include images normal cases, Choroidal Neovascularization (CNV), Drusen, Diabetic Macular Edema (DME). Through experimentation simulation, improves accuracy 1.3%, yielding higher than other SOTA models. We also performed ablation study demonstrate effectiveness method. achieved overall average 99.50% 99.65% datasets.

Language: Английский

---

3DVT: Hyperspectral Image Classification Using 3D Dilated Convolution and Mean Transformer

Photonics, Journal Year: 2025, Volume and Issue: 12(2), P. 146 - 146

Published: Feb. 11, 2025

Hyperspectral imaging and laser technology both rely on different wavelengths of light to analyze the characteristics materials, revealing their composition, state, or structure through precise spectral data. In hyperspectral image (HSI) classification tasks, limited number labeled samples lack feature extraction diversity often lead suboptimal performance. Furthermore, traditional convolutional neural networks (CNNs) primarily focus local features in data, neglecting long-range dependencies global context. To address these challenges, this paper proposes a novel model that combines CNNs with an average pooling Vision Transformer (ViT) for classification. The utilizes three-dimensional dilated convolution two-dimensional extract multi-scale spatial–spectral features, while ViT was employed capture Unlike encoder, which uses linear projection, our replaces it projection. This change enhances compensates encoder’s limitations extraction. hybrid approach effectively strengths dependency handling capabilities Transformers, significantly improving overall performance tasks. Additionally, proposed method holds promise fiber spectra, where high precision analysis are crucial distinguishing between characteristics. Experimental results demonstrate CNN-Transformer substantially improves accuracy three benchmark datasets. accuracies achieved public datasets—IP, PU, SV—were 99.35%, 99.31%, 99.66%, respectively. These advancements offer potential benefits wide range applications, including high-performance optical sensing, medicine, environmental monitoring, accurate is essential development advanced systems fields such as medicine technology.

Language: Английский

---

SAM-OCTA: Prompting segment-anything for OCTA image segmentation

Biomedical Signal Processing and Control, Journal Year: 2025, Volume and Issue: 106, P. 107698 - 107698

Published: Feb. 21, 2025

Language: Английский

---

Redesigning Deep Neural Networks: Bridging Game Theory and Statistical Physics

Published: Jan. 1, 2025

Language: Английский

---

Spectral Transition Evaluation and Heatmap Extraction for Deep Learning Classifiers

Lecture notes in electrical engineering, Journal Year: 2025, Volume and Issue: unknown, P. 439 - 450

Published: Jan. 1, 2025

Language: Английский

---