UNet++: Redesigning Skip Connections to Exploit Multiscale Features in Image Segmentation

IEEE Transactions on Medical Imaging, Год журнала: 2019, Номер 39(6), С. 1856 - 1867

Опубликована: Дек. 13, 2019

The state-of-the-art models for medical image segmentation are variants of U-Net and fully convolutional networks (FCN). Despite their success, these have two limitations: (1) optimal depth is apriori unknown, requiring extensive architecture search or inefficient ensemble varying depths; (2) skip connections impose an unnecessarily restrictive fusion scheme, forcing aggregation only at the same-scale feature maps encoder decoder sub-networks. To overcome limitations, we propose UNet++, a new neural semantic instance segmentation, by alleviating unknown network with efficient U-Nets depths, which partially share co-learn simultaneously using deep supervision; redesigning to aggregate features scales sub-networks, leading highly flexible scheme; (3) devising pruning scheme accelerate inference speed UNet++. We evaluated UNet++ six different datasets, covering multiple imaging modalities such as computed tomography (CT), magnetic resonance (MRI), electron microscopy (EM), demonstrating that consistently outperforms baseline task across datasets backbone architectures; enhances quality varying-size objects-an improvement over fixed-depth U-Net; Mask RCNN++ (Mask R-CNN design) original segmentation; (4) pruned achieve significant speedup while showing modest performance degradation. Our implementation pre-trained available https://github.com/MrGiovanni/UNetPlusPlus.

Язык: Английский

---

Dense Nested Attention Network for Infrared Small Target Detection

IEEE Transactions on Image Processing, Год журнала: 2022, Номер 32, С. 1745 - 1758

Опубликована: Авг. 22, 2022

Single-frame infrared small target (SIRST) detection aims at separating targets from clutter backgrounds. With the advances of deep learning, CNN-based methods have yielded promising results in generic object due to their powerful modeling capability. However, existing cannot be directly applied since pooling layers networks could lead loss layers. To handle this problem, we propose a dense nested attention network (DNA-Net) paper. Specifically, design interactive module (DNIM) achieve progressive interaction among high-level and low-level features. repetitive DNIM, information can maintained. Based on further cascaded channel spatial (CSAM) adaptively enhance multi-level our DNA-Net, contextual well incorporated fully exploited by fusion enhancement. Moreover, develop an dataset (namely, NUDT-SIRST) set evaluation metrics conduct comprehensive performance evaluation. Experiments both public self-developed datasets demonstrate effectiveness method. Compared other state-of-the-art methods, method achieves better terms probability ( Pd ), false-alarm rate Fa intersection union IoU ).

Язык: Английский

---

Machine Learning Methods for Small Data Challenges in Molecular Science

Chemical Reviews, Год журнала: 2023, Номер 123(13), С. 8736 - 8780

Опубликована: Июнь 29, 2023

Small data are often used in scientific and engineering research due to the presence of various constraints, such as time, cost, ethics, privacy, security, technical limitations acquisition. However, big have been focus for past decade, small their challenges received little attention, even though they technically more severe machine learning (ML) deep (DL) studies. Overall, challenge is compounded by issues, diversity, imputation, noise, imbalance, high-dimensionality. Fortunately, current era characterized technological breakthroughs ML, DL, artificial intelligence (AI), which enable data-driven discovery, many advanced ML DL technologies developed inadvertently provided solutions problems. As a result, significant progress has made decade. In this review, we summarize analyze several emerging potential molecular science, including chemical biological sciences. We review both basic algorithms, linear regression, logistic regression (LR),

Язык: Английский

---

BRAVE-NET: Fully Automated Arterial Brain Vessel Segmentation in Patients With Cerebrovascular Disease

Frontiers in Artificial Intelligence, Год журнала: 2020, Номер 3

Опубликована: Сен. 25, 2020

Introduction: Arterial brain vessel assessment is crucial for the diagnostic process in patients with cerebrovascular disease. Non-invasive neuroimaging techniques, such as time-of-flight (TOF) magnetic resonance angiography (MRA) imaging are applied clinical routine to depict arteries. They are, however, only visually assessed. Fully automated segmentation integrated into could facilitate time-critical diagnosis of abnormalities and might identification valuable biomarkers events. In present work, we developed validated a new deep learning model segmentation, coined BRAVE-NET, on large aggregated dataset diseases. Methods: BRAVE-NET multiscale 3-D convolutional neural network (CNN) 264 from three different studies enrolling A context path, dually capturing high- low-resolution volumes, supervision were implemented. The was compared baseline Unet variants paths supervision, respectively. models using high-quality manual labels ground truth. Next precision recall, performance assessed quantitatively by Dice coefficient (DSC); average Hausdorff distance (AVD); 95-percentile (95HD); via visual qualitative rating. Results: surpassed other arterial DSC = 0.931, AVD 0.165, 95HD 29.153. also most resistant toward false labelings revealed analysis. improvement primarily attributed integration multiscaling path lesser extent architectural component. Discussion: We state-of-the-art tailored pathology. provide an extensive experimental validation encompassing variability disease external set healthy volunteers. framework provides technological foundation improving workflow can serve biomarker extraction tool

Язык: Английский

---

R2U++: a multiscale recurrent residual U-Net with dense skip connections for medical image segmentation

Neural Computing and Applications, Год журнала: 2022, Номер 34(20), С. 17723 - 17739

Опубликована: Июнь 3, 2022

Abstract U-Net is a widely adopted neural network in the domain of medical image segmentation. Despite its quick embracement by imaging community, performance suffers on complicated datasets. The problem can be ascribed to simple feature extracting blocks: encoder/decoder, and semantic gap between encoder decoder. Variants (such as R2U-Net) have been proposed address blocks making deeper, but it does not deal with problem. On other hand, another variant UNET++ deals introducing dense skip connections has extraction blocks. To overcome these issues, we propose new based segmentation architecture R2U++. In architecture, adapted changes from vanilla are: (1) plain convolutional backbone replaced deeper recurrent residual convolution block. increased field view aids crucial features for which proven improvement overall network. (2) decoder reduced pathways. These pathways accumulate coming multiple scales apply concatenation accordingly. modified embedded multi-depth models, an ensemble outputs taken varying depths improves foreground objects appearing at various images. R2U++ evaluated four distinct modalities: electron microscopy, X-rays, fundus, computed tomography. average gain achieved IoU score 1.5 ± 0.37% dice 0.9 0.33% over UNET++, whereas, 4.21 2.72 3.47 1.89 R2U-Net across different

Язык: Английский

---

CAA-YOLO: Combined-Attention-Augmented YOLO for Infrared Ocean Ships Detection

Sensors, Год журнала: 2022, Номер 22(10), С. 3782 - 3782

Опубликована: Май 16, 2022

Infrared ocean ships detection still faces great challenges due to the low signal-to-noise ratio and spatial resolution resulting in a severe lack of texture details for small infrared targets, as well distribution extremely multiscale ships. In this paper, we propose CAA-YOLO alleviate problems. study, highlight preserve features apply high-resolution feature layer (P2) better use shallow location information. order suppress noise P2 further enhance extraction capability, introduce TA module into backbone. Moreover, design new fusion method capture long-range contextual information targets combined attention mechanism ability while suppressing interference caused by layers. We conduct detailed study algorithm based on marine dataset verify effectiveness our algorithm, which AP AR increase 5.63% 9.01%, respectively, mAP increases 3.4% compared that YOLOv5.

Язык: Английский

---

Data-Driven Deep Supervision for Medical Image Segmentation

IEEE Transactions on Medical Imaging, Год журнала: 2022, Номер 41(6), С. 1560 - 1574

Опубликована: Янв. 14, 2022

Medical image segmentation plays a vital role in disease diagnosis and analysis. However, data-dependent difficulties such as low contrast, noisy background, complicated objects of interest render the problem challenging. These diminish dense prediction make it tough for known approaches to explore data-specific attributes robust feature extraction. In this paper, we study medical by focusing on extraction achieve improved prediction. We propose new deep convolutional neural network (CNN), which exploits specific input datasets utilize supervision enhanced particular, strategically locate deploy auxiliary supervision, matching object perceptive field (OPF) (which define compute) with layer-wise effective receptive fields (LERF) network. This helps model pay close attention some distinct data dependent features, might otherwise 'ignore' during training. Further, better target localization refined prediction, densely decoded networks (DDN), selectively introducing additional connections (the xmlns:xlink="http://www.w3.org/1999/xlink">'crutch' connections). Using five public (two retinal vessel, melanoma, optic disc/cup, spleen segmentation) two in-house (lymph node fungus segmentation), verify effectiveness our proposed approach 2D 3D segmentation.

Язык: Английский

---

Deep learning applications in breast cancer histopathological imaging: diagnosis, treatment, and prognosis

Breast Cancer Research, Год журнала: 2024, Номер 26(1)

Опубликована: Сен. 20, 2024

Язык: Английский

---

DenseNets Reloaded: Paradigm Shift Beyond ResNets and ViTs

Lecture notes in computer science, Год журнала: 2024, Номер unknown, С. 395 - 415

Опубликована: Окт. 28, 2024

Язык: Английский

---

CMM-Net: Contextual multi-scale multi-level network for efficient biomedical image segmentation

Scientific Reports, Год журнала: 2021, Номер 11(1)

Опубликована: Май 13, 2021

Abstract Medical image segmentation of tissue abnormalities, key organs, or blood vascular system is great significance for any computerized diagnostic system. However, automatic in medical analysis a challenging task since it requires sophisticated knowledge the target organ anatomy. This paper develops an end-to-end deep learning method called Contextual Multi-Scale Multi-Level Network (CMM-Net). The main idea to fuse global contextual features multiple spatial scales at every contracting convolutional network level U-Net. Also, we re-exploit dilated convolution module that enables expansion receptive field with different rates depending on size feature maps throughout networks. In addition, augmented testing scheme referred as Inversion Recovery (IR) which uses logical “OR” and “AND” operators developed. proposed evaluated three imaging datasets, namely ISIC 2017 skin lesions from dermoscopy images, DRIVE retinal vessels fundus BraTS 2018 brain gliomas MR scans. experimental results showed superior state-of-the-art performance overall dice similarity coefficients 85.78%, 80.27%, 88.96% lesions, vessels, tumors, respectively. CMM-Net inherently general could be efficiently applied robust tool various segmentations.

Язык: Английский

---