Quantitative CT radiomics-based models for prediction of haematoma expansion and poor functional outcome in primary intracerebral haemorrhage

Abstract

Objectives: To test radiomics-based features extracted from noncontrast CT of patients with spontaneous intracerebral haemorrhage for prediction of haematoma expansion and poor functional outcome and compare them with radiological signs and clinical factors.

Materials and methods: Seven hundred fifty-four radiomics-based features were extracted from 1732 scans derived from the TICH-2 multicentre clinical trial. Features were harmonised and a correlation-based feature selection was applied. Different elastic-net parameterisations were tested to assess the predictive performance of the selected radiomics-based features using grid optimisation. For comparison, the same procedure was run using radiological signs and clinical factors separately. Models trained with radiomics-based features combined with radiological signs or clinical factors were tested. Predictive performance was evaluated using the area under the receiver operating characteristic curve (AUC) score.

Results: The optimal radiomics-based model showed an AUC of 0.693 for haematoma expansion and an AUC of 0.783 for poor functional outcome. Models with radiological signs alone yielded substantial reductions in sensitivity. Combining radiomics-based features and radiological signs did not provide any improvement over radiomics-based features alone. Models with clinical factors had similar performance compared to using radiomics-based features, albeit with low sensitivity for haematoma expansion. Performance of radiomics-based features was boosted by incorporating clinical factors, with time from onset to scan and age being the most important contributors for haematoma expansion and poor functional outcome prediction, respectively.

Conclusion: Radiomics-based features perform better than radiological signs and similarly to clinical factors on the prediction of haematoma expansion and poor functional outcome. Moreover, combining radiomics-based features with clinical factors improves their performance.

Key points: • Linear models based on CT radiomics-based features perform better than radiological signs on the prediction of haematoma expansion and poor functional outcome in the context of intracerebral haemorrhage. • Linear models based on CT radiomics-based features perform similarly to clinical factors known to be good predictors. However, combining these clinical factors with radiomics-based features increases their predictive performance.

Keywords: Cerebral parenchymal hemorrhage; Linear models; Predictive medicine; Radiomics.

Fig. 1

Study inclusion flowchart

Fig. 2

Feature extraction process flowchart. NCCT scans and their annotations are resampled to 1mm isotropic. Shape features are extracted from the resampled annotations and intensity and texture features are extracted from the resampled original and filtered images. This set of features, together with ultra-early haematoma growth are harmonised and the final set of uncorrelated features is then computed using a correlation-based filtering method. NCCT, noncontrast computed tomography; LoG, Laplacian of Gaussian

Fig. 3

Training and testing procedure. The training UK data is split into 10 non-overlapping folds and 10 different models are trained for each value of the hyperparameter α, using each fold as validation data once. The model that shows the greatest AUC is selected for testing using the non-UK holdout data

Fig. 4

TSNE visualisations of standardised training and testing radiomics feature vectors for each of the 3 harmonisation batches. Each point represents a feature vector for one subject. The left column corresponds to subject radiomics feature vectors pre-harmonisation and the right column corresponds to subject radiomics feature vectors post-harmonisation

Fig. 5

Threshold analysis for sensitivity, specificity, Youden’s index, F1 score, F0.5 score, and F2 score (left column) and ROC curve (right column) for the five prediction models of haematoma expansion (radiomics, radiological signs, radiomics and signs combined, clinical factors, and radiomics and clinical factors combined). Optimal threshold criterion was maximal Youden’s index

Fig. 6

Threshold analysis for sensitivity, specificity, Youden’s index, F1 score, F0.5 score, and F2 score (left column) and ROC curve (right column) for the three prediction models of poor functional (radiomics, radiological signs, radiomics and signs combined, clinical factors, and radiomics and clinical factors combined). Optimal threshold criterion was maximal Youden’s index