Diffusion-weighted imaging-based radiomics model using automatic machine learning to differentiate cerebral cystic metastases from brain abscesses

Abstract

Objectives: To develop a radiomics model based on diffusion-weighted imaging (DWI) utilizing automated machine learning method to differentiate cerebral cystic metastases from brain abscesses.

Materials and methods: A total of 186 patients with cerebral cystic metastases (n = 98) and brain abscesses (n = 88) from two clinical institutions were retrospectively included. The datasets (129 from institution A) were randomly portioned into separate 75% training and 25% internal testing sets. Radiomics features were extracted from DWI images using two subregions of the lesion (cystic core and solid wall). A thorough image preprocessing method was applied to DWI images to ensure the robustness of radiomics features before feature extraction. Then the Tree-based Pipeline Optimization Tool (TPOT) was utilized to search for the best optimized machine learning pipeline, using a fivefold cross-validation in the training set. The external test set (57 from institution B) was used to evaluate the model's performance.

Results: Seven distinct TPOT models were optimized to distinguish between cerebral cystic metastases and abscesses either based on different features combination or using wavelet transform. The optimal model demonstrated an AUC of 1.00, an accuracy of 0.97, sensitivity of 1.00, and specificity of 0.93 in the internal test set, based on the combination of cystic core and solid wall radiomics signature using wavelet transform. In the external test set, this model reached 1.00 AUC, 0.96 accuracy, 1.00 sensitivity, and 0.93 specificity.

Conclusion: The DWI-based radiomics model established by TPOT exhibits a promising predictive capacity in distinguishing cerebral cystic metastases from abscesses.

Keywords: Brain abscesses; Cerebral cystic metastases; Image normalization; Tree-based optimization tool.

Fig. 1

Representative examples of cerebral cystic metastases and brain abscesses of the capsular stage. Cerebral cystic metastases (a–d) and brain abscesses of the capsular stage (e–h) exhibit similarly rim enhancing on CE-T1WI. a, b A 51-year-old male with cerebral cystic metastasis from lung adenocarcinoma. The core of the lesion presented hyperintensity on DWI. c, d A 56-year-old male patient with cerebral cystic metastasis from esophageal carcinoma. The core of the lesion presented hypointensity on DWI. e, f A 58-year-old male with bacterial brain abscess. The core of the lesion presented hyperintensity on DWI. g, h A 9-year-old male patient with fungal brain abscess. The core of the lesion presented hypointensity on DWI

Fig. 2

Study enrollment flowchart

Fig. 3

The pipeline of image preprocessing. The baseline preprocessing pipeline of four different cases are presented, including skull stripping, resample and histogram normalization

Fig. 4

The workflow of this study

Fig. 5

ROC curves of features’ performance before and after normalization in the internal and external test sets. The images normalized have better feature performance, compared to the original images both in the internal (AUC 1.00 vs. 0.86) and external test set (AUC 0.98 vs. 0.55)

Fig. 6

The ROC curves of seven distinct TPOT models. a The ROC curve of seven distinct TPOT models in the internal test set. b The ROC curve of seven distinct TPOT models in the external test set

Fig. 7

Coefficient of model’s features. a Coefficient of the clinical model’s top ten high-ranking features. b Coefficient of the best radiomics model’s top ten high-ranking radiomics features