文献速递神经胸部MR

新型影像组学特征与深度学习在小样本中鉴别脑转移瘤肺癌病理类型的比较

Novel radiomic features versus deep learning: differentiating brain metastases from pathological lung cancer types in small datasets.
PubMed 文献 · Br J Radiol · 2023-05-02

本研究基于脑转移瘤MRI,纳入75例小细胞肺癌、72例鳞癌和75例腺癌片段,提取Laplacian of Gaussian滤波和二维小波变换的新型影像组学特征,并与EfficientNet和ResNet深度学习分类器比较。影像组学分类器敏感度94.44%、特异度95.33%,ResNet-50敏感度94.29%、特异度94.08%,提示影像组学算法在小样本中可作为辅助诊断工具。

要点

  • 脑转移瘤MRI影像组学特征有助于鉴别肺癌病理亚型。
  • 新型特征来自Laplacian of Gaussian滤波和二维小波变换。
  • 影像组学分类器敏感度94.44%、特异度95.33%。

诊断标准与关键数值

  • 影像组学分类器敏感度94.44%、特异度95.33%。
  • 第二个分类器敏感度87.67%、特异度92.62%。
  • ResNet-50一对其余方法敏感度94.29%、特异度94.08%。
  • EfficientNet-b0一对其余方法敏感度92.86%、特异度93.42%。
  • 一对一方法中两个影像组学模型和两个深度学习模型准确率分别为84.68%、78.37%、92.34%和90.99%。

临床应用与阅片要点

  • 脑转移瘤MRI影像组学分析可作为鉴别肺癌病理亚型的辅助手段。
  • 小样本数据集下影像组学算法与深度学习分类器性能相近,可辅助决策。

局限性

  • 研究基于小样本数据集。
  • 仅纳入脑转移瘤MRI,未涉及其他影像模态或原发灶影像。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

英文原文(摘要)· 点击展开

OBJECTIVE: Accurate diagnosis and early treatment are crucial for survival in patients with brain metastases. This study aims to expand the capability of radiomics-based classification algorithms with novel features and compare results with deep learning-based algorithms to differentiate the subtypes of lung cancer from MRI of metastatic lesions in the brain.

METHODS: This study includes 75 small cell lung carcinoma, 72 squamous cell carcinoma, and 75 adenocarcinoma segments. For the radiomics-based algorithm, novel features from the original Laplacian of Gaussian filtered and two-dimensional wavelet transformed images were extracted, and a new three-stage feature selection algorithm was proposed for feature selection. Two classification methods were applied to images to identify the subtypes of lung cancer. Additionally, EfficientNet and ResNet with transfer learning were used as classifiers to compare the results of the proposed algorithm.

RESULTS: The sensitivity and specificity values of the radiomics-based classifier are 94.44 and 95.33%, and for the second classifier are 87.67% and 92.62%, respectively. Besides, a one-vs-all approach comparison was made utilizing two deep learning-based classifiers; The sensitivity and specificity values of 94.29 and 94.08% were obtained from ResNet-50. Moreover, mentioned metrics for EfficientNet-b0 are 92.86 and 93.42%. Furthermore, the accuracies of two radiomics-based and two deep learning-based models were 84.68%, 78.37%, 92.34%, and 90.99%, respectively for one-vs-one approach.

CONCLUSION: The results suggest that the proposed radiomics-based algorithm is a helpful diagnostic assistant to improve decision-making for treating patients with brain metastases in small datasets.

ADVANCES IN KNOWLEDGE: Firstly, the proposed method of this study extracts novel features from transformations of the original images, such as wavelet and Laplacian of Gaussian filter for the first time in literature. Secondly, this is the first study that investigates the classification performance of the shallow and deep learning approaches to identify subtypes of lung cancer.

来源:Br J Radiol,查看原文。版权归原作者所有。

引用

  1. Tulum. Novel radiomic features versus deep learning: differentiating brain metastases from pathological lung cancer types in small datasets. Br J Radiol. 2023. PMID: 37129296. doi:10.1259/bjr.20220841

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。