Distributed training of foundation models for ophthalmic diagnosis DOI Creative Commons
Sina Gholami,

Fatema-E- Jannat,

Atalie C. Thompson

и другие.

Communications Engineering, Год журнала: 2025, Номер 4(1)

Опубликована: Янв. 22, 2025

Vision impairment affects nearly 2.2 billion people globally, and half of these cases could be prevented with early diagnosis intervention—underscoring the urgent need for reliable scalable detection methods conditions like diabetic retinopathy age-related macular degeneration. Here we propose a distributed deep learning framework that integrates self-supervised domain-adaptive federated to enhance eye diseases from optical coherence tomography images. We employed self-supervised, mask-based pre-training strategy develop robust foundation encoder. This encoder was trained on seven datasets, compared its performance under local, centralized, settings. Our results show methods—both centralized federated—improved area curve by at least 10% local models. Additionally, incorporating domain adaptation into further boosted generalization across different populations imaging conditions. approach supports collaborative model development without data sharing, providing scalable, privacy-preserving solution effective retinal disease screening in diverse clinical Sina Gholami co-authors use improve multi-class classification Their preserves privacy ensuring robust, significant gains.

Язык: Английский

Early detection of red palm weevil infestations using deep learning classification of acoustic signals DOI
Wadii Boulila, Ayyub Alzahem, Anis Koubâa

и другие.

Computers and Electronics in Agriculture, Год журнала: 2023, Номер 212, С. 108154 - 108154

Опубликована: Авг. 28, 2023

Язык: Английский

Процитировано

20
EResNet‐SVM: an overfitting‐relieved deep learning model for recognition of plant diseases and pests DOI
Haitao Xiong, Juan Li,

Tiewei Wang

и другие.

Journal of the Science of Food and Agriculture, Год журнала: 2024, Номер 104(10), С. 6018 - 6034

Опубликована: Март 14, 2024

Abstract BACKGROUND The accurate recognition and early warning for plant diseases pests are a prerequisite of intelligent prevention control pests. As result the phenotype similarity hazarded after occur, as well interference external environment, traditional deep learning models often face overfitting problem in pests, which leads to not only slow convergence speed network, but also low accuracy. RESULTS Motivated by above problems, present study proposes model EResNet‐support vector machine (SVM) alleviate classification First, feature extraction capability is improved increasing layers convolutional neural network. Second, order‐reduced modules embedded sparsely activated function introduced reduce complexity overfitting. Finally, classifier fused SVM fully connected transforms original non‐linear into linear high‐dimensional space further improve accuracy ablation experiments demonstrate that structure can effectively experimental results typical show proposed EResNet‐SVM has 99.30% test eight conditions (seven one normal), 5.90% higher than ResNet18. Compared with classic AlexNet, GoogLeNet, Xception, SqueezeNet DenseNet201 models, 5.10%, 7%, 8.10%, 6.20% 1.90%, respectively. testing 6 insect 100%, 3.90% ResNet18 model. CONCLUSION This research provides useful references alleviating learning, theoretical technical support detection © 2024 Society Chemical Industry.

Язык: Английский

Процитировано

9
LeafSpotNet: A deep learning framework for detecting leaf spot disease in jasmine plants DOI Creative Commons

V. Shwetha,

Arnav Bhagwat,

Vijaya Laxmi

и другие.

Artificial Intelligence in Agriculture, Год журнала: 2024, Номер 12, С. 1 - 18

Опубликована: Март 11, 2024

Leaf blight spot disease, caused by bacteria and fungi, poses a threat to plant health, leading leaf discoloration diminished agricultural yield. In response, we present MobileNetV3-based classifier designed for the Jasmine plant, leveraging lightweight Convolutional Neural Networks (CNNs) accurately identify disease stages. The model integrates depth-wise convolution layers max pool enhanced feature extraction, focusing on crucial low-level features indicative of disease. Through preprocessing techniques, including data augmentation with Conditional GAN Particle Swarm Optimization selection, achieves robust performance. Evaluation curated datasets demonstrates an outstanding 97% training accuracy, highlighting its efficacy. Real-world testing diverse conditions, such as extreme camera angles varied lighting, attests model's resilience, yielding test accuracies between 94% 96%. dataset's tailored design CNN-based classification ensures result reliability. Importantly, classification, marked fast computation time reduced size, positions it efficient solution real-time applications. This comprehensive approach underscores proposed classifier's significance in addressing challenges commercial crops.

Язык: Английский

Процитировано

8
FSM-YOLO: Apple leaf disease detection network based on adaptive feature capture and spatial context awareness DOI
Chunman Yan,

Kangyi Yang

Digital Signal Processing, Год журнала: 2024, Номер 155, С. 104770 - 104770

Опубликована: Сен. 7, 2024

Язык: Английский

Процитировано

7
Distributed training of foundation models for ophthalmic diagnosis DOI Creative Commons
Sina Gholami,

Fatema-E- Jannat,

Atalie C. Thompson

и другие.

Communications Engineering, Год журнала: 2025, Номер 4(1)

Опубликована: Янв. 22, 2025

Vision impairment affects nearly 2.2 billion people globally, and half of these cases could be prevented with early diagnosis intervention—underscoring the urgent need for reliable scalable detection methods conditions like diabetic retinopathy age-related macular degeneration. Here we propose a distributed deep learning framework that integrates self-supervised domain-adaptive federated to enhance eye diseases from optical coherence tomography images. We employed self-supervised, mask-based pre-training strategy develop robust foundation encoder. This encoder was trained on seven datasets, compared its performance under local, centralized, settings. Our results show methods—both centralized federated—improved area curve by at least 10% local models. Additionally, incorporating domain adaptation into further boosted generalization across different populations imaging conditions. approach supports collaborative model development without data sharing, providing scalable, privacy-preserving solution effective retinal disease screening in diverse clinical Sina Gholami co-authors use improve multi-class classification Their preserves privacy ensuring robust, significant gains.

Язык: Английский

Процитировано

1