PaperlayerView exact revision on arXiv

Source-linked AI summary

COVID-19 Chest CT Image Segmentation -- A Deep Convolutional Neural Network Solution

Qingsen Yan, Bo Wang, Dong Gong, Chuan Luo, Wei Zhao, Jianhu Shen, Qinfeng Shi, Shuo Jin, Liang Zhang, Zheng You

arXiv:2004.10987v2eess.IVcs.CVcs.LG

TL;DR

COVID-19 CT segmentation is difficult because infection boundaries resemble lung and chest-wall backgrounds, while RT-PCR sensitivity may be limited. COVID-SegNet addresses this with feature variation and progressive atrous pooling, achieving improved infection segmentation and strong lung-segmentation results.

  • Problem

    COVID-19 infection boundaries are difficult to distinguish in chest CT images, while RT-PCR sensitivity may be insufficient for detection.

  • Method

    COVID-SegNet uses feature variation blocks for adaptive global contrast enhancement and Progressive Atrous Spatial Pyramid Pooling for multiscale contextual features.

  • Results

    COVID-SegNet achieves the best results across COVID-19 segmentation metrics, improving Dice over UNet by 3.8%, while lung segmentation reaches an average Dice of 0.987.

  • Takeaways & Limitations

    The study supports using a tailored three-dimensional convolutional network to segment COVID-19 infections and lungs from chest CT images.

Abstract

from arXiv · show

A novel coronavirus disease 2019 (COVID-19) was detected and has spread rapidly across various countries around the world since the end of the year 2019, Computed Tomography (CT) images have been used as a crucial alternative to the time-consuming RT-PCR test. However, pure manual segmentation of CT images faces a serious challenge with the increase of suspected cases, resulting in urgent requirements for accurate and automatic segmentation of COVID-19 infections. Unfortunately, since the imaging characteristics of the COVID-19 infection are diverse and similar to the backgrounds, existing medical image segmentation methods cannot achieve satisfactory performance. In this work, we try to establish a new deep convolutional neural network tailored for segmenting the chest CT images with COVID-19 infections. We firstly maintain a large and new chest CT image dataset consisting of 165,667 annotated chest CT images from 861 patients with confirmed COVID-19. Inspired by the observation that the boundary of the infected lung can be enhanced by adjusting the global intensity, in the proposed deep CNN, we introduce a feature variation block which adaptively adjusts the global properties of the features for segmenting COVID-19 infection. The proposed FV block can enhance the capability of feature representation effectively and adaptively for diverse cases. We fuse features at different scales by proposing Progressive Atrous Spatial Pyramid Pooling to handle the sophisticated infection areas with diverse appearance and shapes. We conducted experiments on the data collected in China and Germany and show that the proposed deep CNN can produce impressive performance effectively.

I. INTRODUCTION

COVID-19 CT lesions are difficult to distinguish from surrounding lung and chest-wall structures, motivating automatic segmentation. The paper proposes COVID-SegNet with feature variation, progressive multi-scale aggregation, and a large annotated dataset.

  • CT imaging can complement RT-PCR for COVID-19 screening, but manually contouring lung lesions is tedious and labor-intensive as cases increase.CT commonly shows bilateral patchy shadows or ground-glass opacity, while RT-PCR sensitivity may be limited by specimen quality and viral material.
  • COVID-19 infection regions have diverse appearances and difficult boundaries, including ground-glass opacities and consolidation across the lung lobes.The boundaries can be hard to distinguish from the chest wall, requiring adjustment of window breadth and locations during annotation.
  • COVID-SegNet is a 3D convolutional network for segmenting both COVID-19 infection regions and the entire lung from chest CT images.The network uses feature variation to adaptively enhance feature contrast and intensity for confusing boundaries.
  • The Feature Variation block addresses difficulty distinguishing COVID-19 pneumonia from lung structures by adaptively adjusting global feature properties.Its design is motivated by boundary enhancement observed when CT window breadth and locations are adjusted.
  • Progressive Atrous Spatial Pyramid Pooling progressively aggregates multi-scale contextual information to handle infection regions with diverse sizes and shapes.The approach fuses features from different receptive fields for more effective contextual representation.
  • The study maintains an expert-annotated dataset of 21,658 chest CT images from 861 confirmed COVID-19 patients and uses ten German cases to test robustness.The dataset is presented as a contribution supporting training and evaluation of the proposed networks.

II. MATERIALS

The study uses chest CT data from 861 RT-PCR-confirmed COVID-19 patients collected at five Chinese hospitals, with separate training and testing subsets. Expert annotators produced lung and infection-region labels, which a senior radiologist assessed.

  • Chest CT images from 861 RT-PCR-confirmed COVID-19 patients were acquired at five Chinese hospitals.The study received medical-ethics approval and waived further consent.
  • A total of 731 patients’ CT images were randomly selected for training, while images from 130 patients formed the testing set.
  • Six annotators with radiology backgrounds labeled lung and COVID-19 infection areas and their boundaries.A senior radiologist with frontline COVID-19 clinical experience assessed the final annotation quality.
  • The materials section introduces the proposed approach, its Feature Variation and Progressive Atrous Spatial Pyramid Pooling blocks, and the training and implementation details.

A. Network Structure of COVID-SegNet

COVID-SegNet combines an encoder–decoder architecture with Feature Variation blocks and PASPP to extract, refine, and aggregate features for COVID-19 infection segmentation.

  • Network Structure of COVID-SegNet: COVID-SegNet uses an encoder with feature extraction and a decoder that estimates segmentation results.The encoder includes residual and FV-based feature extraction operations, while the decoder produces segmentation outputs.
  • Feature Variation: The FV block combines contrast enhancement, position-sensitive, and identity branches to refine COVID-19 boundary and location features.The branches are fused through concatenation and a 3 × 3 × 3 convolution, followed by residual learning with the input.
  • Position Sensitive Branch: The position-sensitive branch uses spatial attention to emphasize helpful COVID-19 regions and suppress harmful information.Its attention map is generated by two convolutional layers and multiplied element-wise with the input feature.
  • Progressive Atrous Spatial Pyramid Pooling: PASPP aggregates semantic information progressively across features with different receptive fields produced by atrous convolutions.It receives FV-refined features and uses different dilation rates to obtain multi-scale representations.
  • Contrast Enhancement Branch: The contrast enhancement branch learns a global parameter from the input feature and applies a shared channel weighting to enhance feature contrast.Global average pooling and convolution produce the parameter, whose expanded weights are applied element-wise to the feature map.

2) Position Sensitive Branch:

The position-sensitive branch generates an attention map that highlights useful COVID-19 segmentation features while suppressing harmful information.

  • Position Sensitive Branch: The position-sensitive branch uses two 3×3×3 convolutional layers followed by ReLU and sigmoid functions to calculate an attention map.The resulting map has the same size as the input feature, with values in the range [0, 1].
  • Position Sensitive Branch: The branch obtains its output by multiplying the input feature element-wise with the attention map.This operation focuses the representation on features useful for COVID-19 infection segmentation.

C. Progressive Atrous Spatial Pyramid Pooling

PASPP addresses infections with varied sizes by using atrous convolutions at multiple receptive fields and progressively aggregating their outputs.

  • Atrous Spatial Pyramid Pooling: Atrous convolution increases the receptive field while preserving feature-map size, and its dilation rate controls the receptive-field extent.The filter size is K, the dilation rate is d, and w[k] denotes the k-th filter parameter.
  • Atrous Spatial Pyramid Pooling: ASPP applies atrous convolutions with different dilation rates to generate and concatenate features containing information at various scales.These multi-scale outputs provide samples of the input with different scale information.
  • Progressive Atrous Spatial Pyramid Pooling: The PASPP structure assembles two residual branches in parallel and progressively blends their outputs after summing 1×1×1 convolution results.The block produces the output feature Fpout from the assembled information.
  • Progressive Atrous Spatial Pyramid Pooling: PASPP is used because COVID-19 infection regions often have very different sizes and therefore require features with different receptive fields.The block progressively fuses these features rather than only concatenating them.
  • Progressive Atrous Spatial Pyramid Pooling: PASPP first applies four parallel 1×1×1 convolutions, reducing each branch to one quarter of the input channel count.The resulting branch features are processed as inputs to the multi-scale aggregation structure.
  • Progressive Atrous Spatial Pyramid Pooling: Outputs from adjacent atrous-convolution branches are summed and added to residual-branch outputs before subsequent processing.The branch outputs are then progressively aggregated using adjacent features in parallel.

A. Dataset

The study uses annotated chest CT images from confirmed COVID-19 patients and evaluates segmentation with Dice, sensitivity, and precision.

  • The dataset contains 21,658 annotated chest CT images from 861 confirmed COVID-19 patients.Images from 731 patients are used for training and those from 130 patients for testing.
  • Segmentation performance is evaluated using Dice similarity coefficient, sensitivity, and precision.Dice measures similarity between predicted and reference score maps.

C. Implementation Details

COVID-SegNet uses residual, feature-variation, and progressive atrous pyramid components in a 3D convolutional framework, trained and compared under shared settings. Qualitative tests cover difficult domestic cases involving low contrast, chest-wall adjacency, and lung segmentation.

  • Implementation Details: The network uses residual blocks, FV blocks, PASPP blocks, downsampling, deconvolution, and a final softmax layer.Convolutional layers generally use 3 × 3 × 3 kernels with batch normalization and ReLU.
  • Implementation Details: Training uses randomly cropped 128 × 128 × 64 patches, Adam optimization, an initial learning rate of 1e−4, and batch size 2.Models are trained from scratch using six Nvidia TITAN RTX GPUs.
  • Implementation Details: COVID-SegNet is compared with FCN, UNet, VNet, and UNet++ using 3D convolution, the same training data, and shared settings.The network performs both lung and COVID-19 segmentation.
  • Qualitative Results on the Domestic Datasets: On low-contrast and difficult domestic cases, the proposed method produces more coherent infection boundaries and avoids scattered artifacts.The FV blocks adaptively enhance global feature contrast, while PASPP further improves performance.
  • Qualitative Results on the Domestic Datasets: For lung segmentation, the compared methods generally approach manual annotations, while UNet++ can miss boundaries and VNet can produce nonsmooth margins.The proposed method is shown with a 3D lung surface rendering.

2) Qualitative Results on the Germany Data:

Ten German cases are used to assess generalization for COVID-19 infection and lung segmentation. The proposed method is reported to handle challenging intensity similarity and marked lung regions more effectively than the compared methods.

  • Ten cases from Brainlab Co. Ltd. in Germany are used to test COVID-19 infection and lung segmentation.
  • On German COVID-19 scans with infection intensity similar to lung tissue, competing methods produce poor or over-segmented results, whereas the proposed method resembles manual annotation.The proposed method also segments small infection regions in the 3D rendering.
  • For German lung segmentation, most methods produce distinct outlines, while the proposed method shows stronger segmentation in regions marked by red arrows.The authors associate these results with the FV and PASPP blocks.

3) Quantitative Results:

Quantitative evaluation reports the proposed method as best across COVID-19 segmentation metrics and gives strong lung-segmentation averages. Ablation analyses examine the FV block’s contrast and position components and PASPP variants.

  • Quantitative Results: The proposed method achieves the best results across all reported COVID-19 segmentation metrics and improves Dice over UNet by 3.8%.
  • Quantitative Results: For lung segmentation, the proposed method achieves average Dice 0.987, sensitivity 0.986, and precision 0.990.It surpasses state-of-the-art methods in precision despite already strong existing performance.
  • Study on the FV block: The ablation study evaluates CEB, PSB, FV, CAB, ASPP, and ResASPP variants against a four-layer UNet baseline.These comparisons assess global contrast enhancement, position-sensitive features, and PASPP design choices.
  • Study on the FV block: Combining global contrast and local position information in the FV block improves segmentation performance, while the reported comparisons demonstrate FV effectiveness.PSB focuses on feature positions useful for detecting and segmenting COVID-19 infection regions.

2) Study on the PASPP block:

The study evaluates PASPP and shows that progressive multi-scale feature fusion improves COVID-19 segmentation, while the complete network performs best on both lung and infection segmentation.

  • 2) Study on the PASPP block:: Progressively fusing features in PASPP is effective for COVID-19 segmentation.The block uses atrous convolutions with different dilation rates and progressive concatenations.
  • 2) Study on the PASPP block:: The proposed network achieves the best performance on both lung segmentation and COVID-19 segmentation.
  • 2) Study on the PASPP block:: COVID-SegNet combines feature variation and progressive ASPP blocks to highlight infection boundaries and positions in three-dimensional chest CT images.The model was evaluated using manually annotated CT data from 861 confirmed COVID-19 patients collected at five hospitals.
Loading 2004.10987v2…