Application of machine learning based on radiomics in the discrimination of intracranial germ cell tumours

Abstract

Background: Among germ cell tumours, germinomas are extremely sensitive to radiotherapy and chemotherapy. Histological diagnosis is important for clinical treatment decisions. This study aimed to identify germinomas and non-germinomatous germ cell tumours (NGGCTs) using radiomics-based machine learning (ML).

Methods: The present retrospective study comprised 141 patients diagnosed with intracranial germ cell tumours (ICGCTs), 71 germinomas, and 70 NGGCTs. Radiomics features were quantitatively extracted from magnetic resonance imaging (MRI) sequences, including T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), T2 Fluid-Attenuated Inversion Recovery (T2-FLAIR), diffusion weighted imaging (DWI) (b=1,000), apparent diffusion coefficient (ADC) images, and contrast-enhanced T1WI. Based on the combination of three feature selection methods and three classification methods, the optimal model was screened out from the internal test set. A combined model of clinical-multi-sequence radiomics was ultimately created by combining with statistically significant clinical features. The performance of the models was evaluated using the area under the curve (AUC), accuracy, sensitivity, specificity, and F1-score.

Results: The combination of the least absolute shrinkage and selection operator (LASSO) and logistic regression (LR) yielded the optimal diagnostic performance in the multi-sequence radiomics model, as evidenced by an AUC value of 0.823 in the internal and 0.804 in the external test set. In the combined model, the AUC values of the internal and external tests were 0.838 and 0.809, respectively. The DeLong test revealed no significant difference between multi-sequence radiomics and the combined model, indicating that the inclusion of clinical characteristics did not significantly improve diagnostic accuracy.

Conclusions: ML based on radiomics may provide a non-invasive approach for the clinical differentiation of intracranial germinomas and NGGCTs.

Keywords: Machine learning (ML); germinomas; magnetic resonance imaging (MRI); non-germinomatous germ cell tumours (NGGCTs); radiomics.

Figure 1

Flowchart for patient inclusion in the two hospital districts. MRI, magnetic resonance imaging; NGGCTs, non-germinomatous germ cell tumours.

Figure 2

MRI images of four representative patients with intracranial germ-cell tumours showed germ-cell tumour sellar area, pineal area and NGGCTs sellar area, pineal area from top to bottom. Different histological types have the same imaging findings: cystic change, uneven and obvious enhancement, partial diffusion limitation on DWI. ADC, apparent diffusion coefficient; DWI, diffusion weighted imaging; MRI, magnetic resonance imaging; NGGCTs, non-germinomatous germ cell tumours.

Figure 3

The flowchart illustrates the process of radiomics, which encompasses the stages of image acquisition, feature extraction, feature selection, and model analysis. ADC, apparent diffusion coefficient; DWI, diffusion weighted imaging; LASSO, least absolute shrinkage and selection operator; LOG, Laplacian of Gaussian; MR, magnetic resonance; MRI, magnetic resonance imaging; RFE, Recursive Feature Elimination; ROC, receiver operating characteristic.

Figure 4

The line plots and bar contrast plots are used to visually compare model performance under and between different feature choices in the internal test set. AUC, area under the curve; LASSO, least absolute shrinkage and selection operator; LR, logistic regression; mRMR, minimum Redundancy Maximum Relevance; RF, random forest; RFE, Recursive Feature Elimination; SVM, support vector machine.

Figure 5

The ROC, decision curves, and calibration curves for the radiomic, clinical, and combined models are presented. The images in panels (A-C) illustrate the internal test set performance, while panels (D-F) depict the external test set performance. AUC, area under the curve; ROC, receiver operating characteristic.

Figure 6

SHAP. By calculating the SHAP value of each feature in the combined model under the training set, the importance of each feature in the model prediction is explained, and the SHAP bar plot (A), swarm plots (B), and heat maps (C) were visually displayed. The colour red represents positive contributions, while blue represents negative contributions. SHAP, Shapley Additive exPlanations.

Figure 7

Permutation importance of box diagram. Based on the box-plot drawn from 100 times of feature permutation evaluation, the ordinate reflects the level of feature importance.