ABSTRACT
Objective:To distinguish lung metastases of different origin by constructing a classification model according to CT radiomics features.Methods:A total of 226 patients with lung metastases of gastric cancer, breast cancer and kidney cancer attending Chongqing Red Cross Hospital from January 2015 to July 2020, with a total of 402 metastases, were randomly divided into a training cohort (training set, 136 patients, 280 metastases) and a validation cohort (validation set, 90 patients, 122 metastases) by the hold-out method. In addition, 68 patients with lung metastases (138 lung metastases in total) attending Chongqing Red Cross Hospital from August 2020 to April 2022 were matched as an external test cohort (test set). Region of interest segmentation was performed by two experienced radiologists independently and manually without clinical information to construct the model by using LASSO screening for the best radiomic features. Support vector machine (SVM) and random forest (RF) were selected to build dichotomous and trichotomous models respectively. The receiver operating characteristic curve was used to evaluate the classification efficiency of both models.Results:There were no statistically significant differences in age ( t=-0.06, P=0.534), gender ( χ2<0.01, P=0.961) and number of lung metastases ( χ2=0.71, P=0.703) between the validation and test sets. A total of 792 radiomic features were extracted, 703 of which had good agreement (intraclass correlation coefficient≥0.75), while 89 features being excluded for having poor agreement (intraclass correlation coefficient<0.75). The dichotomous model (SVM) screened 28 (lung metastases from gastric cancer vs. lung metastases from breast cancer), 25 (lung metastases from gastric cancer vs. lung metastases from kidney cancer) and 34 (lung metastases from kidney cancer vs. lung metastases from breast cancer) features, respectively; the trichotomous model (RF) screened 20 features (three types of lung metastases), in which Short Run Emphasis and Inverse Variance were significantly higher in lung metastases from kidney cancer than in the other two types, correlation was higher in lung metastases from gastric cancer than in the other two types, and there was no significant difference in the sphericity of the three lung metastases. For the dichotomous model, in the validation set, the area under the curve (AUC) of the 28 features selected to distinguish gastric cancer lung metastases from breast cancer lung metastases was 0.81, the AUC of the 25 features distinguishing gastric cancer lung metastases from kidney cancer lung metastases was 0.86, and the AUC of the 34 features distinguishing kidney cancer lung metastases from breast cancer lung metastases was 0.92, and the AUCs of the test set were 0.80, 0.79 and 0.86 respectively. For the trichotomous model, the AUC for predicting lung metastases from gastric cancer, breast cancer and kidney cancer in the validation set were 0.85, 0.82 and 0.91 respectively, and both macroscopic and microscopic AUC were 0.85; In the test set, the AUC for predicting lung metastases from gastric cancer, breast cancer, and kidney cancer were 0.77, 0.86 and 0.84 respectively, and both macroscopic and microscopic AUC were 0.81. Conclusion:The SVM and RF models based on CT radiomic features are helpful in distinguishing lung metastases derived from gastric cancer, breast cancer and kidney cancer.
ABSTRACT
Precisely segmented lung fields restrict the region-of-interest from which radiological patterns are searched, and is thus an indispensable prerequisite step in any chest radiographic CADx system. Recently, a number of deep learning-based approaches have been proposed to implement this step. However, deep learning has its own limitations and cannot be used in resource-constrained settings. Medical systems generally have limited RAM, computational power, storage, and no GPUs. They are thus not always suited for running deep learning-based models. Shallow learning-based models with appropriately selected features give comparable performance but with modest resources. The present paper thus proposes a shallow learning-based method that makes use of 40 radiomic features to segment lung fields from chest radiographs. A distance regularized level set evolution (DRLSE) method along with other post-processing steps are used to refine its output. The proposed method is trained and tested using publicly available JSRT dataset. The testing results indicate that the performance of the proposed method is comparable to the state-of-the-art deep learning-based lung field segmentation (LFS) methods and better than other LFS methods.