Medical Image Segmentation via Cascaded Attention Decoding

2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), Год журнала: 2023, Номер unknown, С. 6211 - 6220

Опубликована: Янв. 1, 2023

Transformers have shown great promise in medical image segmentation due to their ability capture long-range dependencies through self-attention. However, they lack the learn local (contextual) relations among pixels. Previous works try overcome this problem by embedding convolutional layers either encoder or decoder modules of transformers thus ending up sometimes with inconsistent features. To address issue, we propose a novel attention-based decoder, namely CASCaded Attention DEcoder (CASCADE), which leverages multi-scale features hierarchical vision transformers. CASCADE consists i) an attention gate fuses skip connections and ii) module that enhances context suppressing background information. We use multi-stage feature loss aggregation framework faster convergence better performance. Our experiments demonstrate significantly outperform state-of-the-art CNN- transformer-based approaches, obtaining 5.07% 6.16% improvements DICE mIoU scores, respectively. opens new ways designing decoders.

Язык: Английский

---

High-Resolution Iterative Feedback Network for Camouflaged Object Detection

Proceedings of the AAAI Conference on Artificial Intelligence, Год журнала: 2023, Номер 37(1), С. 881 - 889

Опубликована: Июнь 26, 2023

Spotting camouflaged objects that are visually assimilated into the background is tricky for both object detection algorithms and humans who usually confused or cheated by perfectly intrinsic similarities between foreground surroundings. To tackle this challenge, we aim to extract high-resolution texture details avoid detail degradation causes blurred vision in edges boundaries. We introduce a novel HitNet refine low-resolution representations features an iterative feedback manner, essentially global loop-based connection among multi-scale resolutions. design better feature flow corruption caused recurrent path, strategy proposed impose more constraints on each connection. Extensive experiments four challenging datasets demonstrate our breaks performance bottleneck achieves significant improvements compared with 29 state-of-the-art methods. In addition, address data scarcity scenarios, provide application example convert salient objects, thereby generating training samples from diverse datasets. Code will be made publicly available.

Язык: Английский

---

MTANet: Multi-Task Attention Network for Automatic Medical Image Segmentation and Classification

IEEE Transactions on Medical Imaging, Год журнала: 2023, Номер 43(2), С. 674 - 685

Опубликована: Сен. 19, 2023

Medical image segmentation and classification are two of the most key steps in computer-aided clinical diagnosis. The region interest were usually segmented a proper manner to extract useful features for further disease classification. However, these methods computationally complex time-consuming. In this paper, we proposed one-stage multi-task attention network (MTANet) which efficiently classifies objects an while generating high-quality mask each medical object. A reverse addition module was designed task fusion areas global map boundary cues high-resolution features, bottleneck used feature fusion. We evaluated performance MTANet with CNN-based transformer-based architectures across three imaging modalities different tasks: CVC-ClinicDB dataset polyp segmentation, ISIC-2018 skin lesion our private ultrasound liver tumor Our model outperformed state-of-the-art models on all datasets superior 25 radiologists

Язык: Английский

---

G-CASCADE: Efficient Cascaded Graph Convolutional Decoding for 2D Medical Image Segmentation

2022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), Год журнала: 2024, Номер unknown, С. 7713 - 7722

Опубликована: Янв. 3, 2024

In this paper, we are the first to propose a new graph convolution-based decoder namely, Cascaded Graph Convolutional Attention Decoder (G-CASCADE), for 2D medical image segmentation. G-CASCADE progressively refines multi-stage feature maps generated by hierarchical transformer encoders with an efficient convolution block. The encoder utilizes self-attention mechanism capture long-range dependencies, while preserving information due global receptive fields of Rigorous evaluations our multiple on five segmentation tasks (i.e., Abdomen organs, Cardiac Polyp lesions, Skin and Retinal vessels) show that model outperforms other state-of-the-art (SOTA) methods. We also demonstrate achieves better DICE scores than SOTA CASCADE 80.8% fewer parameters 82.3% FLOPs. Our can easily be used general-purpose semantic tasks. implementation found at: https://github.com/SLDGroup/G-CASCADE.

Язык: Английский

---

EMCAD: Efficient Multi-Scale Convolutional Attention Decoding for Medical Image Segmentation

2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Год журнала: 2024, Номер 40, С. 11769 - 11779

Опубликована: Июнь 16, 2024

Язык: Английский

---

Lightweight medical image segmentation network with multi-scale feature-guided fusion

Computers in Biology and Medicine, Год журнала: 2024, Номер 182, С. 109204 - 109204

Опубликована: Окт. 3, 2024

Язык: Английский

---

META-Unet: Multi-Scale Efficient Transformer Attention Unet for Fast and High-Accuracy Polyp Segmentation

IEEE Transactions on Automation Science and Engineering, Год журнала: 2023, Номер 21(3), С. 4117 - 4128

Опубликована: Июль 12, 2023

Polyp segmentation plays an important role in preventing Colorectal cancer. Although Vision Transformer has been widely introduced medical image to compensate the limitations of traditional CNN modeling global context, its shortcomings learning fine-detailed features and heavy computation cost also hinder application challenging polyp due various shapes sizes polyps, low-intensity contrast between polyps surrounding tissues, inherent real-time requirement. In this paper, we propose a multi-scale efficient transformer attention (META) mechanism for fast high-accuracy segmentation, where blocks are employed generate element-wise attentions adaptive feature fusion famous U-shape encoder-decoder architecture. Specifically, our META includes two branches capture long-term dependencies, which implemented via with different resolutions. The local branch is used relatively smaller transform under lower resolution, while high-resolution attention. final poly results progressively integrated based on each layer decoder. Extensive experiments conducted four datasets (CVC-ClinicDB, Endoscenestill, Kvasir-SEG ETIS-Larib) demonstrate advantages, consistently outperforming competitors. While using ResNet34 as backbones, it can achieve 85.78% IoU 92.03% Dice, 88.99% 93.85% 86.42% 91.86% Dice respectively CVC-ClinicDB, Kvasir-SEG, speed 98 FPS at input size $3 \times 512 512$ NVIDIA GeForce RTX 3090 card. code available https://github.com/szuzzb/META-Unet. Note Practitioners —Automatic crucial step recognition diagnostic colonoscopy, usually require both performance. This article proposes novel method, namely META-Unet, by maps effectively efficiently mechanism, faster higher-accuracy segmentation. We evaluate META-Unet public ETIS-Larib). Comprehensive experimental validate outstanding performance better balance accuracy inference speed. proposed potentially be embedded deep frameworks facilitates more computer-aided applications clinical practice.

Язык: Английский

---

Boosting Camouflaged Object Detection with Dual-Task Interactive Transformer

2022 26th International Conference on Pattern Recognition (ICPR), Год журнала: 2022, Номер unknown, С. 140 - 146

Опубликована: Авг. 21, 2022

Camouflaged object detection intends to discover the concealed objects hidden in surroundings. Existing methods follow bio-inspired framework, which first locates and second refines boundary. We argue that discovery of camouflaged depends on recurrent search for The processing makes human tired helpless, but it is just advantage transformer with global ability. Therefore, a dual-task interactive proposed detect both accurate position its detailed boundary feature considered as Query improve detection, meanwhile detection. are fully interacted by multi-head self-attention. Besides, obtain initial feature, transformer-based backbones adopted extract foreground background. object, while minus background Here, can be obtained from blurry region Supervised ground truth, model achieves state-of-the-art performance public datasets. https://github.com/liuzywen/COD

Язык: Английский

---

CoInNet: A Convolution-Involution Network With a Novel Statistical Attention for Automatic Polyp Segmentation

IEEE Transactions on Medical Imaging, Год журнала: 2023, Номер 42(12), С. 3987 - 4000

Опубликована: Сен. 28, 2023

Polyps are very common abnormalities in human gastrointestinal regions. Their early diagnosis may help reducing the risk of colorectal cancer. Vision-based computer-aided diagnostic systems automatically identify polyp regions to assist surgeons their removal. Due varying shape, color, size, texture, and unclear boundaries, segmentation images is a challenging problem. Existing deep learning models mostly rely on convolutional neural networks that have certain limitations diversity visual patterns at different spatial locations. Further, they fail capture inter-feature dependencies. Vision transformer also been deployed for due powerful global feature extraction capabilities. But too supplemented by convolution layers contextual local information. In present paper, model CoInNet proposed with novel mechanism leverages strengths involution operations learns highlight considering relationship between maps through statistical attention unit. To further aid network an anomaly boundary approximation module introduced uses recursively fed fusion refine results. It indeed remarkable even tiny-sized polyps only 0.01% image area can be precisely segmented CoInNet. crucial clinical applications, as small easily overlooked manual examination voluminous size wireless capsule endoscopy videos. outperforms thirteen state-of-the-art methods five benchmark datasets.

Язык: Английский

---

Uncertainty-Aware Hierarchical Aggregation Network for Medical Image Segmentation

IEEE Transactions on Circuits and Systems for Video Technology, Год журнала: 2024, Номер 34(8), С. 7440 - 7453

Опубликована: Фев. 26, 2024

Medical image segmentation is an essential process to assist clinics with computer-aided diagnosis and treatment. Recently, a large amount of convolutional neural network (CNN)-based methods have been rapidly developed achieved remarkable performances in several different medical tasks. However, the same type infected region or lesions often has diversity scales, making it challenging task achieve accurate segmentation. In this paper, we present novel Uncertainty-aware Hierarchical Aggregation Network, namely UHA-Net, for segmentation, which can fully make utilization cross-level multi-scale features handle scale variations. Specifically, propose hierarchical feature fusion (HFF) module aggregate high-level features, used produce global map coarse localization segmented target. Then, uncertainty-induced (UCF) fuse from adjacent levels, learn knowledge guidance capture contextual information resolutions. Further, aggregation (SAM) presented by using convolution kernels, effectively deal At last, formulate unified framework simultaneously inter-layer discriminability representations intra-layer leading results. We carry out experiments on three tasks, results demonstrate that our UHA-Net outperforms state-of-the-art methods. Our implementation code maps will be publicly at https://github.com/taozh2017/UHANet.

Язык: Английский

---