Dataset Information

Multi-level dilated residual network for biomedical image segmentation.

ABSTRACT: We propose a novel multi-level dilated residual neural network, an extension of the classical U-Net architecture, for biomedical image segmentation. U-Net is the most popular deep neural architecture for biomedical image segmentation, however, despite being state-of-the-art, the model has a few limitations. In this study, we suggest replacing convolutional blocks of the classical U-Net with multi-level dilated residual blocks, resulting in enhanced learning capability. We also propose to incorporate a non-linear multi-level residual blocks into skip connections to reduce the semantic gap and to restore the information lost when concatenating features from encoder to decoder units. We evaluate the proposed approach on five publicly available biomedical datasets with different imaging modalities, including electron microscopy, magnetic resonance imaging, histopathology, and dermoscopy, each with its own segmentation challenges. The proposed approach consistently outperforms the classical U-Net by 2%, 3%, 6%, 8%, and 14% relative improvements in dice coefficient, respectively for magnetic resonance imaging, dermoscopy, histopathology, cell nuclei microscopy, and electron microscopy modalities. The visual assessments of the segmentation results further show that the proposed approach is robust against outliers and preserves better continuity in boundaries compared to the classical U-Net and its variant, MultiResUNet.

SUBMITTER: Gudhe NR

PROVIDER: S-EPMC8266898 | biostudies-literature | 2021 Jul

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

Multi-level dilated residual network for biomedical image segmentation.

Gudhe Naga Raju NR Behravan Hamid H Sudah Mazen M Okuma Hidemi H Vanninen Ritva R Kosma Veli-Matti VM Mannermaa Arto A

Scientific reports 20210708 1

We propose a novel multi-level dilated residual neural network, an extension of the classical U-Net architecture, for biomedical image segmentation. U-Net is the most popular deep neural architecture for biomedical image segmentation, however, despite being state-of-the-art, the model has a few limitations. In this study, we suggest replacing convolutional blocks of the classical U-Net with multi-level dilated residual blocks, resulting in enhanced learning capability. We also propose to incorpo ...[more]

PMID: 34238940

Similar Datasets

Project description:BackgroundThe segmentation of breast ultrasound (US) images has been a challenging task, mainly due to limited data and the inherent image characteristics involved, such as low contrast and speckle noise. Although convolutional neural network-based (CNN-based) methods have made significant progress over the past decade, they lack the ability to model long-range interactions. Recently, the transformer method has been successfully applied to the tasks of computer vision. It has a strong ability to capture distant interactions. However, most transformer-based methods with excellent performance rely on pre-training on large datasets, making it infeasible to directly apply them to medical images analysis, especially that of breast US images with limited high-quality labels. Therefore, it is of great significance to find a robust and efficient transformer-based method for use on small breast US image datasets.MethodsWe developed a dilated transformer (DT) method which mainly uses the proposed residual axial attention layers to build encoder blocks and the introduced dilation module (DM) to further increase the receptive field. We evaluated the proposed method on 2 breast US image datasets using the 5-fold cross-validation method. Dataset A was a public dataset with 562 images, while dataset B was a private dataset with 878 images. Ground truth (GT) was delineated by 2 radiologists with more than 5 years of experience. The evaluation was followed by related ablation experiments.ResultsThe DT was found to be comparable with the state-of-the-art (SOTA) CNN-based method and outperformed the related transformer-based method, medical transformer (MT), on both datasets. Especially on dataset B, the DT outperformed the MT on metrics of Jaccard index (JI) and Dice similarity coefficient (DSC) by 2.67% and 4.68%, respectively. Meanwhile, when compared with Unet, the DT improved JI and DSC by 4.89% and 4.66%, respectively. Moreover, the results of the ablation experiments showed that each add-on part of the DT is important and contributes to the segmentation accuracy.ConclusionsThe proposed transformer-based method could achieve advanced segmentation performance on different small breast US image datasets without pretraining.

Project description:BackgroundThe precise identification of the position and form of a tumor mass can improve early diagnosis and treatment. However, due to the complicated tumor categories and varying sizes and forms, the segregation of brain gliomas and their internal sub-regions is still very challenging. This study sought to design a new deep-learning network based on three-dimensional (3D) U-Net to address its shortcomings in brain tumor segmentation (BraTS) tasks.MethodsWe developed a 3D dilated multi-scale residual attention U-Net (DMRA-U-Net) model for magnetic resonance imaging (MRI) BraTS. It used dilated convolution residual (DCR) modules to better process shallow features, multi-scale convolution residual (MCR) modules in the bottom encoding path to create richer and more comprehensive feature expression while reducing overall information loss or blurring, and a channel attention (CA) module between the encoding and decoding paths to address the problem of retrieving and preserving important features during the processing of deep feature maps.ResultsThe BraTS 2018-2021 datasets served as the training and evaluation datasets for this study. Further, the proposed architecture was assessed using metrics such as the dice similarity coefficient (DSC), Hausdorff distance (HD), and sensitivity (Sens). The DMRA U-Net model segments the whole tumor (WT), and the tumor core (TC), and the enhancing tumor (ET) regions of brain tumors. Using the suggested architecture, the DSCs were 0.9012, 0.8867, and 0.8813, the HDs were 28.86, 13.34, and 10.88 mm, and the Sens was 0.9429, 0.9452, and 0.9303 for the WT, TC, and ET regions, respectively. Compared to the traditional 3D U-Net, the DSC of the DMRA U-Net increased by 4.5%, 2.5%, and 0.8%, the HD of the DMRA U-Net decreased by 21.83, 16.42, and 10.00, the Sens of the DMRA U-Net increased by 0.4%, 0.7%, and 1.4% for the WT, TC, and ET regions, respectively. Further, the results of the statistical comparison of the performance indicators revealed that our model performed well generally in the segmentation of the WT, TC, and ET regions.ConclusionsWe developed a promising tumor segmentation model. Our solution is open sourced and is available at: https://github.com/Gold3nk/dmra-unet.

Project description:BackgroundImage segmentation is an important step during the processing of medical images. For example, for the computer aid diagnostic systems for lung cancer image analysis, the segmented regions of tumors would help doctors in early diagnosis to determine timely and appropriate treatment possibilities and thereby improve the survival rate of the patients. However, general clinical routines of manual segmentation for large number of medical images are very difficult and time consuming, which is the challenge we aim to tackle using our proposed method.MethodsA novel image segmentation method with evolutionary learning technique named Group Theoretic Particle Swarm Optimization is proposed. It can tackle multi-level thresholding optimization problem during the segmentation process and rebuild the search paradigm according to the solid mathematical foundation of symmetric group from four designable aspects, which are particle encoding, solution landscape, neighborhood movement and swarm topology, respectively. The Kapur's entropy of multi-level thresholds is assessed as the objective function.ResultsIn contrast to those conventional metaheuristics methods for lung cancer image segmentation, this newly presented method generates the best performance result among them. Experimental results show that its Kapur's entropy has the value of 9.07, which is 16% higher than the worst case. Computational time is acceptable at the cost of 173.730 seconds, average level of evaluation metrics [Kappa, Precision, Recall, F1-measure, intersection over union (IoU) and receiver operating characteristic (ROC)] is over 90%, and search process of multi-level threshold combination would finally converge in the later phase of iterations after 700. The ablation study indicates that all components are significant to the contributions of our proposed method.ConclusionsGroup Theoretic Particle Swarm Optimization for multi-level threshold segmentation is an efficient way to split a medical image into distinct regions and extract tumor tissues regions from the background. It maintains the balanced relationship between diversification and intensification during the search process and helps clinicians to make the diagnosis more accurately. Our proposed method processes potential medical value and clinical meanings.

Dataset Information

Multi-level dilated residual network for biomedical image segmentation.

Publications

Multi-level dilated residual network for biomedical image segmentation.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets