Skip to main content
QUICK REVIEW

[论文解读] Lung segmentation with NASNet-Large-Decoder Net

Youshan Zhang|arXiv (Cornell University)|Mar 18, 2023
Lung Cancer Diagnosis and Treatment被引用 5
一句话总结

本文提出了一种新颖的肺部分割模型,采用NASNet-Large作为编码器并结合自定义解码器架构,Dice分数达到0.92,优于当前最先进方法。该方法集成了后处理层以去除无关区域,显著提升了在RSNA肺炎检测数据集的胸片图像上的分割精度。

ABSTRACT

Lung cancer has emerged as a severe disease that threatens human life and health. The precise segmentation of lung regions is a crucial prerequisite for localizing tumors, which can provide accurate information for lung image analysis. In this work, we first propose a lung image segmentation model using the NASNet-Large as an encoder and then followed by a decoder architecture, which is one of the most commonly used architectures in deep learning for image segmentation. The proposed NASNet-Large-decoder architecture can extract high-level information and expand the feature map to recover the segmentation map. To further improve the segmentation results, we propose a post-processing layer to remove the irrelevant portion of the segmentation map. Experimental results show that an accurate segmentation model with 0.92 dice scores outperforms state-of-the-art performance.

研究动机与目标

  • 开发一种用于早期肺癌检测的高精度肺部分割深度学习模型。
  • 利用NASNet-Large强大的特征提取能力进行医学图像分割。
  • 通过自定义解码器和后处理层提升分割性能。
  • 解决由图像强度变化和解剖结构模糊性带来的肺部分割挑战。
  • 建立一种强大且可迁移的医学图像分割框架,超越肺部分割任务。

提出的方法

  • 使用NASNet-Large作为编码器,从胸部X光(CXR)图像中提取高层级、分层的特征。
  • 采用四模块解码器结构,结合上采样、卷积层、批归一化和ReLU激活函数,重建分割图。
  • 应用最终的Softmax层进行像素级肺部分割预测。
  • 引入后处理层以过滤掉非肺部区域(如软组织或伪影),提升掩码精度。
  • 在GPU上使用RSNA肺炎检测数据集进行端到端训练,批量大小为4,训练100个周期,优化器为Adam。
  • 采用Dice和IoU作为评估指标,量化分割性能。
Fig. 1 : The architecture of NASNet-Large segmentation net. The encoder consists of the first 414 layers from the NASNet-Large model. There are four blocks in the decoder, and each block contains Upsamling, Conv+ReLu, and BN layer. (Convolution (Conv), Batch normalization (BN), Rectified linear unit
Fig. 1 : The architecture of NASNet-Large segmentation net. The encoder consists of the first 414 layers from the NASNet-Large model. There are four blocks in the decoder, and each block contains Upsamling, Conv+ReLu, and BN layer. (Convolution (Conv), Batch normalization (BN), Rectified linear unit

实验结果

研究问题

  • RQ1NASNet-Large能否作为低剂量CXR图像中肺部分割的有效编码器?
  • RQ2所提出的解码器架构在捕捉细微肺部边界方面,与标准U-Net或SegNet相比表现如何?
  • RQ3后处理层在多大程度上减少了肺部分割中的假阳性?
  • RQ4NASNet-Large-Decoder Net是否在RSNA数据集上实现了当前最先进性能?
  • RQ5该模型对图像强度变化(如过度曝光或曝光不足)的鲁棒性如何?

主要发现

  • 所提出的NASNet-Large-Decoder Net实现了0.92的Dice分数和0.87的IoU,优于SegNet(0.87 Dice)、U-Net(0.88 Dice)和DeepLab(0.89 Dice)。
  • 经后处理后的模型性能最高,Dice分数达0.92,IoU为0.87,证实了后处理层的有效性。
  • 视觉对比显示,预测的分割图即使在复杂的解剖区域也与真实掩码高度吻合。
  • 在极端亮度或暗度的图像中观察到失败案例,表明模型在分布外图像上的泛化能力有限。
  • 该模型在标准指标上表现优异,表明在肺部区域定位方面具有高精度和高召回率。
  • 该架构具备可迁移性,可适配于肺部分割以外的其他医学图像分割任务。
Fig. 2 : The post-processing of prediction result. The left one is the prediction result from the proposed net, and the right one is the post-processing result using the proposed post-processing layer. The red box is an irrelevant feature.
Fig. 2 : The post-processing of prediction result. The left one is the prediction result from the proposed net, and the right one is the post-processing result using the proposed post-processing layer. The red box is an irrelevant feature.

更好的研究,从现在开始

从阅读论文到最终审阅,大幅缩短您的研究时间。

无需绑定信用卡

本解读由 AI 生成,并经人工编辑审核。