PaperlayerView exact revision on arXiv

Source-linked AI summary

MiniSeg: An Extremely Minimum Network for Efficient COVID-19 Segmentation

Yu Qiu, Yun Liu, Shijie Li, Jing Xu

arXiv:2004.09750v3cs.CV

TL;DR

Limited COVID-19 training data and the computational demands of traditional segmentation hinder efficient CT-based screening. The paper proposes MiniSeg, an 83K-parameter lightweight network using AHSP multi-scale learning, and builds a benchmark for comparison. MiniSeg achieves favorable performance against previous state-of-the-art methods with high efficiency, while showing DSC above 0.7 for most slices and median DSC above 0.8 across lesion counts on COVID-19-P20.

  • Problem

    Limited COVID-19 training data can cause overfitting in traditional data-hungry models, while computationally intensive methods hinder fast training, testing, and deployment.

  • Method

    MiniSeg is an 83K-parameter lightweight segmentation network incorporating the AHSP module for multi-scale learning and accompanied by a comprehensive COVID-19 segmentation benchmark.

  • Results

    MiniSeg performs favorably against previous state-of-the-art segmentation methods with high efficiency; on COVID-19-P20, DSC exceeds 0.7 for most slices and median DSC exceeds 0.8 regardless of lesion counts.

  • Takeaways & Limitations

    MiniSeg provides an efficient COVID-19 infected-area segmentation approach designed for limited training data and practical screening-system deployment.

Abstract

from arXiv · show

The rapid spread of the new pandemic, i.e., COVID-19, has severely threatened global health. Deep-learning-based computer-aided screening, e.g., COVID-19 infected CT area segmentation, has attracted much attention. However, the publicly available COVID-19 training data are limited, easily causing overfitting for traditional deep learning methods that are usually data-hungry with millions of parameters. On the other hand, fast training/testing and low computational cost are also necessary for quick deployment and development of COVID-19 screening systems, but traditional deep learning methods are usually computationally intensive. To address the above problems, we propose MiniSeg, a lightweight deep learning model for efficient COVID-19 segmentation. Compared with traditional segmentation methods, MiniSeg has several significant strengths: i) it only has 83K parameters and is thus not easy to overfit; ii) it has high computational efficiency and is thus convenient for practical deployment; iii) it can be fast retrained by other users using their private COVID-19 data for further improving performance. In addition, we build a comprehensive COVID-19 segmentation benchmark for comparing MiniSeg to traditional methods.

Introduction

COVID-19 screening needs computer-aided CT interpretation because RT-PCR sensitivity is limited and experienced radiologists are scarce. MiniSeg addresses limited training data and deployment constraints through lightweight multi-scale learning, an 83K-parameter network, and a benchmark for comparison.

  • Motivation: Limited COVID-19 data can cause overfitting in traditional deep-learning models with millions of parameters.Traditional segmentation methods are also computationally intensive, making fast training/testing and low computational load important for screening-system deployment.
  • Motivation: Lightweight networks are proposed to reduce overfitting risk and computational cost while preserving accurate segmentation through multi-scale learning.The paper identifies effective multi-scale learning as a way to improve segmentation accuracy under parameter and efficiency constraints.
  • Method: AHSP performs lightweight multi-scale learning using dilated depthwise separable convolutions, feature pooling, hierarchical fusion, and attention-guided feature merging.The attention mechanism highlights essential information and filters noisy information in radiography images.
  • Contributions: 83K parameters make MiniSeg two orders of magnitude smaller than traditional image segmentation methods.The authors state that the limited COVID-19 data can therefore be sufficient for training MiniSeg.
  • Contributions: The paper builds a comprehensive COVID-19 segmentation benchmark to compare MiniSeg extensively with previous methods.The benchmark is presented as a resource for evaluating segmentation approaches under the paper’s target setting.

Related Work

Prior segmentation methods emphasize multi-scale representation and accuracy but often overlook model size and inference speed. Lightweight architectures reduce computational complexity, motivating COVID-19 segmentation methods that improve accuracy and speed within a lightweight setting.

  • Multi-scale segmentation: Encoder-decoder networks, dilated-convolution ASPP modules, pyramid pooling, context encoding, and non-local operations are used to learn multi-scale or global-context representations.These approaches primarily target segmentation accuracy without considering model size and inference speed.
  • Lightweight networks: Lightweight networks reduce parameters and improve efficiency through convolutional factorization, including depthwise separable, grouped, and channel-shuffle operations.Examples include Flattened Model, Inception, Xception, MobileNets, and ShuffleNets.
  • Lightweight segmentation: Efficient semantic-segmentation research provides relevant designs for reducing computational cost in segmentation networks.The paper positions its COVID-19 goal as achieving higher accuracy and faster speed by enhancing multi-scale learning in a lightweight setting.
  • COVID-19 screening: COVID-19 screening studies commonly classify chest X-rays or CT slices, while Fan et al. proposed CT infected-area segmentation but remained within the category of previous segmentation methods.The paper presents this prior segmentation approach as suboptimal relative to its lightweight objective.

Methodology

MiniSeg combines the AHSP module with an encoder-decoder architecture to learn multi-scale features efficiently for COVID-19 CT segmentation. Its design uses dilated depthwise separable convolutions, hierarchical aggregation, spatial attention, and lightweight feature fusion.

  • Attentive Hierarchical Spatial Pyramid Module: AHSP builds multi-scale features with parallel dilated depthwise separable convolutions and pooling, then aggregates them hierarchically.Large and small dilation rates capture contextual and local information, respectively.
  • Attentive Hierarchical Spatial Pyramid Module: Spatial attention weights each scale to emphasize target structures and suppress irrelevant information during multi-scale learning.The attention map is generated by a pointwise convolution and sigmoid activation before element-wise multiplication with each feature map.
  • Attentive Hierarchical Spatial Pyramid Module: AHSP uses grouped pointwise convolution to fuse hierarchical feature maps, reducing Eqn. 8 parameters from C′2 to C′2/K.This design separately fuses feature branches and reduces the number of convolution parameters by K times compared with vanilla pointwise convolution.
  • MiniSeg Network Architecture: MiniSeg uses an encoder-decoder structure in which the encoder learns multi-scale representations and the decoder aggregates representations across levels.The decoder combines high-level semantic features with low-level fine details for infected-area prediction.
  • MiniSeg Network Architecture: The encoder uses AHSP blocks, nested skip pathways, and four downsampled scales, while the decoder produces dense two-class predictions with deep supervision.The four encoder scales are 1/2, 1/4, 1/8, and 1/16; the prediction classes are background and COVID-19.

Experimental Setup

MiniSeg is evaluated on four open-access COVID-19 CT segmentation datasets using standardized training, augmentation, cross-validation, and medical-imaging metrics.

  • Training: All methods use PyTorch, Adam optimization, 80 training epochs, batch size 5, and identical training settings for fair comparison.The initial learning rate is 1e-3 with weight decay of 1e-4 and a poly learning-rate policy.
  • Datasets: Four open-access COVID-19 CT segmentation datasets are used to evaluate MiniSeg and the baselines.The datasets are renamed COVID-19-CT100, COVID-19-P9, COVID-19-P20, and COVID-19-P1110.
  • Evaluation Protocol: Standard cropping and random flipping augment the training data, while 5-fold cross-validation reduces statistically significant differences in performance evaluation.
  • Metrics: Performance is measured using mIoU, SEN, SPC, DSC, and HD, where higher values are better except for HD.mIoU, SEN, SPC, and DSC range from 0 to 1, whereas lower HD indicates better segmentation accuracy.

Ablation Studies

Ablation experiments evaluate MiniSeg’s main components and design choices, with results supporting the effectiveness of its complete architecture.

  • Main Components: Adding multi-scale branches, attentive hierarchical fusion, the two-path encoder, and channel splitting produces the final MiniSeg configuration.The ablation sequence begins with a single-branch module and progressively adds the main architectural components.
  • Main Components: The ablation studies conclude that MiniSeg’s main components are all effective for COVID-19 segmentation.
  • Design Choices: The default setting achieves the best overall performance across tested activation, decoder, supervision, convolution, and feature-fusion alternatives.The evaluated choices include PReLU versus ReLU, decoder removal, deep-supervision removal, AHSP substitutions, kernel-size changes, and FFM replacements.

Comparison with State-of-the-art Methods

MiniSeg is compared with 31 previous state-of-the-art segmentation methods through computational, numerical, qualitative, and statistical analyses across COVID-19 datasets.

  • Benchmark: The benchmark contains 31 previous state-of-the-art image segmentation methods for extensive comparison with MiniSeg.
  • Efficiency: MiniSeg has extremely small parameter and FLOP counts and is much faster than competing methods.The comparison covers parameters, FLOPs, and speed.
  • Quantitative Results: MiniSeg achieves the best or close-to-best performance across all metrics and datasets, with limited exceptions for SPC and SEN.It is slightly below the best SPC method on COVID-19-CT100 and COVID-19-P9, and does not achieve the best SEN on COVID-19-P1110.
  • Qualitative Results: Visual comparisons against eight state-of-the-art methods further indicate that MiniSeg outperforms the baseline methods.Figure 4 encodes true positives in red, false negatives in green, and false positives in blue.
  • Statistical Analysis: On COVID-19-P20, most CT slices have DSC above 0.7 regardless of infected area, while median DSC exceeds 0.8 regardless of lesion count.

Conclusion

MiniSeg addresses limited COVID-19 training data and deployment efficiency through an extremely small network with AHSP-based multi-scale learning and a comprehensive benchmark.

  • Conclusion: MiniSeg uses the AHSP module to support accurate and efficient COVID-19 infected-area segmentation under an extremely small network-size constraint.
  • Conclusion: The benchmark compares MiniSeg extensively with previous state-of-the-art methods and is intended to support future COVID-19 segmentation research.
  • Conclusion: Comparison with state-of-the-art methods demonstrates that MiniSeg combines best performance with high efficiency for practical deployment.
Loading 2004.09750v3…