Non-Local U-Nets for Biomedical Image Segmentation

Wang, Zhengyang; Zou, Na; Shen, Dinggang; Ji, Shuiwang

doi:10.1609/aaai.v34i04.6100

Cited by 138 publications

(72 citation statements)

References 20 publications

(27 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…As a compromise between 3D feature requirement and limited computation resource, some methods use the 3D patching strategy, which clips a small cube as an input sample. 20 , 26 Such 3D methods can extract local shape information very well but raise another problem: the global structure information is lost in clipped patches. This can be ignored in some applications like tumor detection in pathology slices but is very important for the lesions that have a global position tendency in the body, like IAs.…”

Section: Introductionmentioning

confidence: 99%

Toward human intervention-free clinical diagnosis of intracranial aneurysm via deep neural network

Qiao

Tian

et al. 2021

Patterns

View full text Add to dashboard Cite

Summary Intracranial aneurysm (IA) is an enormous threat to human health, which often results in nontraumatic subarachnoid hemorrhage or dismal prognosis. Diagnosing IAs on commonly used computed tomographic angiography (CTA) examinations remains laborious and time consuming, leading to error-prone results in clinical practice, especially for small targets. In this study, we propose a fully automatic deep-learning model for IA segmentation that can be applied to CTA images. Our model, called Global Localization-based IA Network (GLIA-Net), can incorporate the global localization prior and generates the fine-grain three-dimensional segmentation. GLIA-Net is trained and evaluated on a big internal dataset (1,338 scans from six institutions) and two external datasets. Evaluations show that our model exhibits good tolerance to different settings and achieves superior performance to other models. A clinical experiment further demonstrates the clinical utility of our technique, which helps radiologists in the diagnosis of IAs.

show abstract

Section: Introductionmentioning

confidence: 99%

Toward human intervention-free clinical diagnosis of intracranial aneurysm via deep neural network

Qiao

Tian

et al. 2021

Patterns

View full text Add to dashboard Cite

show abstract

“…Then, the brain is extracted using a deep 3D-Unet ( Hwang et al, 2019 ) and registered to the MNI-NIH neonatal brain template 1 , with spatial resolution of 0.6 × 0.6 × 0.6 mm 3 (isotropic). Different types of brain tissue (GM, WM, and CSF) are thereafter segmented by an advanced deep non-local 3D-Unet ( Wang Z. et al, 2020 ). Individual templates (MRI + manually segmented tissue labels) utilized for the deep learning approach are then selected evenly across all PMAs.…”

Section: Methodsmentioning

confidence: 99%

“…Our previous work ( Kim et al, 2016 ; Liu et al, 2019 ), despite successfully solving the issue of large morphological alterations, did not fully address the mis-segmentation between WM and CSF due to the confounding intensities between WM and CSF in neonatal MRI ( Li et al, 2019 ) for neonatal brain ( Figure 4 ). To solve this, a non-local 3D-Unet ( Wang Z. et al, 2020 ) is applied for tissue segmentation in our pipeline. Conventional 3D-Unet (like the one used in brain extraction) applies small kernels in convolution operators to scan inputs and extracts local information.…”

Section: Methodsmentioning

confidence: 99%

“…Non-local 3D-Unet was designed by adding a self-attention block, which computes outputs at one position by attending to every position of the input, to both the encoder layer and the decoder layer to aggregate/capture the contextual information. Non-local 3D-Unet has been demonstrated to outperform most state-of-the-art methods in neonatal tissue segmentation ( Wang Z. et al, 2020 ).…”

Section: Methodsmentioning

confidence: 99%

See 1 more Smart Citation

Robust Cortical Thickness Morphometry of Neonatal Brain and Systematic Evaluation Using Multi-Site MRI Datasets

et al. 2021

View full text Add to dashboard Cite

The human brain grows the most dramatically during the perinatal and early post-natal periods, during which pre-term birth or perinatal injury that may alter brain structure and lead to developmental anomalies. Thus, characterizing cortical thickness of developing brains remains an important goal. However, this task is often complicated by inaccurate cortical surface extraction due to small-size brains. Here, we propose a novel complex framework for the reconstruction of neonatal WM and pial surfaces, accounting for large partial volumes due to small-size brains. The proposed approach relies only on T1-weighted images unlike previous T2-weighted image-based approaches while only T1-weighted images are sometimes available under the different clinical/research setting. Deep neural networks are first introduced to the neonatal magnetic resonance imaging (MRI) pipeline to address the mis-segmentation of brain tissues. Furthermore, this pipeline enhances cortical boundary delineation using combined models of the cerebrospinal fluid (CSF)/GM boundary detection with edge gradient information and a new skeletonization of sulcal folding where no CSF voxels are seen due to the limited resolution. We also proposed a systematic evaluation using three independent datasets comprising 736 pre-term and 97 term neonates. Qualitative assessment for reconstructed cortical surfaces shows that 86.9% are rated as accurate across the three site datasets. In addition, our landmark-based evaluation shows that the mean displacement of the cortical surfaces from the true boundaries was less than a voxel size (0.532 ± 0.035 mm). Evaluating the proposed pipeline (namely NEOCIVET 2.0) shows the robustness and reproducibility across different sites and different age-groups. The mean cortical thickness measured positively correlated with post-menstrual age (PMA) at scan (p < 0.0001); Cingulate cortical areas grew the most rapidly whereas the inferior temporal cortex grew the least rapidly. The range of the cortical thickness measured was biologically congruent (1.3 mm at 28 weeks of PMA to 1.8 mm at term equivalent). Cortical thickness measured on T1 MRI using NEOCIVET 2.0 was compared with that on T2 using the established dHCP pipeline. It was difficult to conclude that either T1 or T2 imaging is more ideal to construct cortical surfaces. NEOCIVET 2.0 has been open to the public through CBRAIN (https://mcin-cnim.ca/technology/cbrain/), a web-based platform for processing brain imaging data.

show abstract

“…Roy et al proposed three attention modules modified from the squeeze-and-excitation module and embedded them in different segmentation network to perform multi-organ segmentation [35]. Wang et al proposed global aggregation blocks with a spatial attention mechanism to extract global information of feature maps [36].…”

Section: Attention Mechanisms In Medical Image Tasksmentioning

confidence: 99%

Eso-Net: A Novel 2.5D Segmentation Network With the Multi-Structure Response Filter for the Cancerous Esophagus

Zhou

Huang

et al. 2020

IEEE Access

View full text Add to dashboard Cite

Automatic segmentation of the cancerous esophagus in computed tomography (CT) images is a computer-assisted method that can improve the efficiency of the diagnosis and treatment. Due to the diversity of the cancer stage and location, the anatomical structure of the cancerous esophagus is various. Moreover, the low contrast against surrounding tissues leads to a blurry boundary of the cancerous esophagus. Therefore, existing segmentation networks cannot achieve satisfactory results in automatic segmentation of the cancerous esophagus. In this paper, we propose a novel 2.5D segmentation network named Eso-Net for the cancerous esophagus based on an encoder-decoder architecture. A 3D enhancement filter called Multi-Structure Response Filter (MSRF) is designed to extract 3D structural information as prior knowledge. Furthermore, dilated convolutions and residual connections are employed in the convolutional blocks of Eso-Net for multi-scale feature learning. With 3D structural priors, Prior Attention Modules (PAM) are incorporated into the network to facilitate the transmission of relevant spatial information. The experiments are conducted on the dataset from 30 esophageal cancer patients, and we report an 84.839% dice similarity coefficient, an 85.955% precision, an 83.752% sensitivity, and a 2.583mm Hausdorff distance. The experimental results demonstrate that the proposed method outperforms other existing segmentation networks in this task and can effectively assist doctors in the diagnosis and treatment of esophageal cancer.

show abstract

Non-Local U-Nets for Biomedical Image Segmentation

Cited by 138 publications

References 20 publications

Toward human intervention-free clinical diagnosis of intracranial aneurysm via deep neural network

Toward human intervention-free clinical diagnosis of intracranial aneurysm via deep neural network

Robust Cortical Thickness Morphometry of Neonatal Brain and Systematic Evaluation Using Multi-Site MRI Datasets

Eso-Net: A Novel 2.5D Segmentation Network With the Multi-Structure Response Filter for the Cancerous Esophagus

Contact Info

Product

Resources

About