DRA-Net: Medical image segmentation based on adaptive feature extraction and region-level information fusion

Scientific Reports, Год журнала: 2024, Номер 14(1)

Опубликована: Апрель 27, 2024

Abstract Medical image segmentation is a key task in computer aided diagnosis. In recent years, convolutional neural network (CNN) has made some achievements medical segmentation. However, the convolution operation can only extract features fixed size region at time, which leads to loss of features. The recently popular Transformer global modeling capabilities, but it does not pay enough attention local information and cannot accurately segment edge details target area. Given these issues, we proposed dynamic regional (DRA-Net). Different from above methods, first measures similarity concentrates on different regions. this way, adaptively select scopes for feature extraction, reducing loss. Then, interaction carried out better learn details. At same also design ordered shift multilayer perceptron (MLP) blocks enhance communication within regions, further enhancing network’s ability After several experiments, results indicate that our produces more accurate performance compared other CNN based networks.

Язык: Английский

---

Real-time non-invasive hemoglobin prediction using deep learning-enabled smartphone imaging

BMC Medical Informatics and Decision Making, Год журнала: 2024, Номер 24(1)

Опубликована: Июль 1, 2024

Abstract Background Accurate measurement of hemoglobin concentration is essential for various medical scenarios, including preoperative evaluations and determining blood loss. Traditional invasive methods are inconvenient not suitable rapid, point-of-care testing. Moreover, current models, due to their complex parameters, well-suited mobile settings, which limits the ability conduct frequent rapid This study aims introduce a novel, compact, efficient system that leverages deep learning smartphone technology accurately estimate levels, thereby facilitating accessible assessments. Methods The employed application capture images eye, were subsequently analyzed by neural network trained on data from test data. Specifically, EGE-Unet model was utilized eyelid segmentation, while DHA(C3AE) level prediction. performance evaluated using statistical metrics mean intersection over union (MIOU), F1 Score, accuracy, specificity, sensitivity. model’s assessed absolute error (MAE), mean-square (MSE), root square (RMSE), R^2. Results demonstrated robust in achieving an MIOU 0.78, Score 0.87, accuracy 0.97, specificity 0.98, sensitivity 0.86. prediction yielded promising outcomes with MAE 1.34, MSE 2.85, RMSE 1.69, R^2 0.34. overall size modest at 1.08 M, computational complexity 0.12 FLOPs (G). Conclusions presents groundbreaking approach eliminates need supplementary devices, providing cost-effective, swift, accurate method healthcare professionals enhance treatment planning improve patient care perioperative environments. proposed has potential enable testing can be particularly beneficial settings. Trial Registration clinical trial registered Chinese Clinical Registry (No. ChiCTR2100044138) 20/02/2021.

Язык: Английский

---

Random color transformation for single domain generalized retinal image segmentation

Engineering Applications of Artificial Intelligence, Год журнала: 2024, Номер 136, С. 108907 - 108907

Опубликована: Июль 6, 2024

Язык: Английский

---

Transformer-based cross-modal multi-contrast network for ophthalmic diseases diagnosis

Journal of Applied Biomedicine, Год журнала: 2023, Номер 43(3), С. 507 - 527

Опубликована: Июнь 27, 2023

Язык: Английский

---

Retinal multi-lesion segmentation by reinforcing single-lesion guidance with multi-view learning

Biomedical Signal Processing and Control, Год журнала: 2023, Номер 86, С. 105349 - 105349

Опубликована: Авг. 12, 2023

Язык: Английский

---