Attention UW-Net: A fully connected model for automatic segmentation and annotation of chest X-ray

Computers in Biology and Medicine, Journal Year: 2022, Volume and Issue: 150, P. 106083 - 106083

Published: Sept. 9, 2022

Language: Английский

---

Deep learning for pneumothorax diagnosis: a systematic review and meta-analysis

European Respiratory Review, Journal Year: 2023, Volume and Issue: 32(168), P. 220259 - 220259

Published: June 7, 2023

Background Deep learning (DL), a subset of artificial intelligence (AI), has been applied to pneumothorax diagnosis aid physician diagnosis, but no meta-analysis performed. Methods A search multiple electronic databases through September 2022 was performed identify studies that DL for using imaging. Meta-analysis via hierarchical model calculate the summary area under curve (AUC) and pooled sensitivity specificity both physicians Risk bias assessed modified Prediction Model Study Bias Assessment Tool. Results In 56 63 primary studies, identified from chest radiography. The total AUC 0.97 (95% CI 0.96–0.98) physicians. 84% 79–89%) 85% 73–92%) 96% 94–98%) 98% 95–99%) More than half original (57%) had high risk bias. Conclusions Our review found diagnostic performance models similar physicians, although majority Further AI research is needed.

Language: Английский

---

ConTEXTual Net: A Multimodal Vision-Language Model for Segmentation of Pneumothorax

Deleted Journal, Journal Year: 2024, Volume and Issue: 37(4), P. 1652 - 1663

Published: March 14, 2024

Radiology narrative reports often describe characteristics of a patient's disease, including its location, size, and shape. Motivated by the recent success multimodal learning, we hypothesized that this descriptive text could guide medical image analysis algorithms. We proposed novel vision-language model, ConTEXTual Net, for task pneumothorax segmentation on chest radiographs. Net extracts language features from physician-generated free-form radiology using pre-trained model. then introduced cross-attention between intermediate embeddings an encoder-decoder convolutional neural network to enable guidance analysis. was trained CANDID-PTX dataset consisting 3196 positive cases with annotations 6 different physicians as well clinical reports. Using cross-validation, achieved Dice score 0.716±0.016, which similar degree inter-reader variability (0.712±0.044) computed subset data. It outperformed vision-only models (Swin UNETR: 0.670±0.015, ResNet50 U-Net: 0.677±0.015, GLoRIA: 0.686±0.014, nnUNet 0.694±0.016) competing model (LAVT: 0.706±0.009). Ablation studies confirmed it information led performance gains. Additionally, show certain augmentation methods degraded Net's breaking image-text concordance. also evaluated effects activation functions in module, highlighting efficacy our chosen architectural design.

Language: Английский

---

Fully Convolutional Network for the Semantic Segmentation of Medical Images: A Survey

Diagnostics, Journal Year: 2022, Volume and Issue: 12(11), P. 2765 - 2765

Published: Nov. 11, 2022

There have been major developments in deep learning computer vision since the 2010s. Deep has contributed to a wealth of data medical image processing, and semantic segmentation is salient technique this field. This study retrospectively reviews recent studies on application for tasks imaging proposes potential directions future development, including model augmentation dataset creation. The strengths deficiencies models augmentation, as well their segmentation, were analyzed. Fully convolutional network led creation U-Net its derivatives. Another noteworthy DeepLab. Regarding due low volume images, most focus means increase data. Generative adversarial networks (GAN) via learning. Despite increasing types datasets, there still deficiency datasets specific problems, which should be improved moving forward. Considering ongoing research processing practical clinical problems must addressed ensure that results are properly applied.

Language: Английский

---

Brain tumor segmentation using U-Net in conjunction with EfficientNet

PeerJ Computer Science, Journal Year: 2024, Volume and Issue: 10, P. e1754 - e1754

Published: Jan. 2, 2024

According to the Ten Leading Causes of Death Statistics Report by Ministry Health and Welfare in 2021, cancer ranks as leading cause mortality. Among them, pleomorphic glioblastoma is a common type brain cancer. Brain often occurs with unclear boundaries from normal tissue, necessitating assistance experienced doctors distinguish tumors before surgical resection avoid damaging critical neural structures. In recent years, advancement deep learning (DL) technology, artificial intelligence (AI) plays vital role disease diagnosis, especially field image segmentation. This technology can aid locating measuring tumors, while significantly reducing manpower time costs. Currently, U-Net one primary segmentation techniques. It utilizes skip connections combine high-level low-level feature information, significant improvements accuracy. To further enhance model’s performance, this study explores feasibility using EfficientNetV2 an encoder combination U-net. Experimental results indicate that employing together U-net improve Dice score (loss = 0.0866, accuracy 0.9977, similarity coefficient (DSC) 0.9133).

Language: Английский

---

Automatic Detection of Liver Cancer Using Hybrid Pre-Trained Models

Sensors, Journal Year: 2022, Volume and Issue: 22(14), P. 5429 - 5429

Published: July 20, 2022

Liver cancer is a life-threatening illness and one of the fastest-growing types in world. Consequently, early detection liver leads to lower mortality rates. This work aims build model that will help clinicians determine type tumor when it occurs within region by analyzing images tissue taken from biopsy this tumor. Working stage requires effort, time, accumulated experience must be possessed expert whether malignant needs treatment. Thus, histology can make use obtain an initial diagnosis. study propose deep learning using convolutional neural networks (CNNs), which are able transfer knowledge pre-trained global models decant into single diagnose tumors CT scans. we obtained hybrid capable detecting The best results research reached accuracy 0.995, precision value 0.864, recall 0.979, higher than those other models. It worth noting was tested on limited set data gave good results. used as aid support decisions specialists field save their efforts. In addition, saves effort time incurred treatment specialists, especially during periodic examination campaigns every year.

Language: Английский

---

Image-based classification of wheat spikes by glume pubescence using convolutional neural networks

Frontiers in Plant Science, Journal Year: 2024, Volume and Issue: 14

Published: Jan. 12, 2024

Pubescence is an important phenotypic trait observed in both vegetative and generative plant organs. Pubescent plants demonstrate increased resistance to various environmental stresses such as drought, low temperatures, pests. It serves a significant morphological marker aids selecting stress-resistant cultivars, particularly wheat. In wheat, pubescence visible on leaves, leaf sheath, glumes nodes. Regarding glumes, the presence of plays pivotal role its classification. supplements other spike characteristics, aiding distinguishing between different varieties within wheat species. The determination typically involves visual analysis by expert. However, methods without use binocular loupe tend be subjective, while employing additional equipment labor-intensive. This paper proposes integrated approach determine glume images captured under laboratory conditions using digital camera convolutional neural networks.

Language: Английский

---

Exploring the role of Convolutional Neural Networks (CNN) in dental radiography segmentation: A comprehensive Systematic Literature Review

Engineering Applications of Artificial Intelligence, Journal Year: 2024, Volume and Issue: 133, P. 108510 - 108510

Published: May 11, 2024

Language: Английский

---

Ultrasound image denoising using generative adversarial networks with residual dense connectivity and weighted joint loss

PeerJ Computer Science, Journal Year: 2022, Volume and Issue: 8, P. e873 - e873

Published: Feb. 16, 2022

Ultrasound imaging has been recognized as a powerful tool in clinical diagnosis. Nonetheless, the presence of speckle noise degrades signal-to-noise ultrasound images. Various denoising algorithms cannot fully reduce and retain image features well for imaging. The application deep learning attracted more attention recent years.In article, we propose generative adversarial network with residual dense connectivity weighted joint loss (GAN-RW) to avoid limitations traditional surpass most advanced performance denoising. is based on U-Net architecture which includes four encoder decoder modules. Each modules replaced BN remove noise. discriminator applies series convolutional layers identify differences between translated images desired modality. In training processes, introduce function consisting sum L1 function, binary cross-entropy logit perceptual function.We split experiments into two parts. First, were performed Berkeley segmentation (BSD68) datasets corrupted by simulated speckle. Compared eight existing algorithms, GAN-RW achieved despeckling terms peak ratio (PSNR), structural similarity (SSIM), subjective visual effect. When level was 15, average value increased approximately 3.58% 1.23% PSNR SSIM, respectively. 25, 3.08% 1.84% 50, 1.32% 1.98% Secondly, lymph nodes, foetal head, brachial plexus. proposed method shows higher effect when verifying end, through statistical analysis, highest mean rank Friedman test.

Language: Английский

---

Focal modulation network for lung segmentation in chest X-ray images

TURKISH JOURNAL OF ELECTRICAL ENGINEERING & COMPUTER SCIENCES, Journal Year: 2023, Volume and Issue: 31(6), P. 1006 - 1020

Published: Oct. 7, 2023

Segmentation of lung regions is key importance for the automatic analysis Chest X-Ray (CXR) images, which have a vital role in detection various pulmonary diseases. Precise identification basic prerequisite disease diagnosis and treatment planning. However, achieving precise segmentation poses significant challenges due to factors such as variations anatomical shape size, presence strong edges at rib cage clavicle, overlapping structures resulting from diverse Although commonly considered de-facto standard medical image segmentation, convolutional UNet architecture its variants fall short addressing these challenges, primarily limited ability model long-range dependencies between features. While vision transformers equipped with self-attention mechanisms excel capturing relationships, either coarse-grained global or fine-grained local typically adopted tasks on high-resolution images alleviate quadratic computational cost expense performance loss. This paper introduces focal modulation (FMN-UNet) enhance by effectively aggregating relations reasonable cost. FMN-UNet first encodes CXR via encoder suppress background extract latent feature maps relatively modest resolution. then leverages attention contextual relationships across images. These are convolutionally decoded produce masks. The compared against state-of-the-art methods three public datasets (JSRT, Montgomery, Shenzhen). Experiments each dataset demonstrate superior baselines.

Language: Английский

---