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

We retrospectively included participants recruited prospectively to the TICH-2 international randomised, placebo-controlled clinical trial (ISRCTN93732214) [20]. This trial tested the efficacy and safety of intravenous tranexamic acid in people with acute spontaneous intracerebral haemorrhage presenting within 8 h of symptom onset. Primary outcome was functional status at day 90 measured by modified Rankin scale. Ethical approval for TICH-2 was obtained from the local institutional review board and informed consent was obtained before enrolment, either from the participant or one of their relatives. The rationale, protocol, and inclusion/exclusion criteria for the TICH-2 trial have been reported elsewhere [21]. All 2077 TICH-2 trial participants that had valid baseline and follow-up scans and have been previously reported [13] were eligible for inclusion in this analysis. We excluded 345 participants for our analysis due to clinical and technical reasons (Fig. 1), yielding a total of 1732 participants. Finally, all analyses were performed on a stratified semi-random split of the participants into a training set (N = 1211, 70%) and testing set (N = 521, 30%), forcing both sets to be age- and gender-matched.

Noncontrast CT (NCCT) brain scans were acquired as part of routine clinical care at each of the 124 centres participating in TICH-2 following their local protocol. Baseline scans were acquired before randomisation and follow-up scans were acquired after 24 ± 12 h [21]. There were no restrictions on scanner manufacturer, scanner settings, or slice thickness. Nevertheless, only axial scans were accepted.

The proposed feature extraction process follows the image processing guidelines of the image biomarker standardisation initiative (IBSI) [22, 23]. Firstly, semi-automated volumetric segmentation of intracerebral haemorrhage, perihaematomal oedema, and intraventricular haemorrhage was performed from each baseline and follow-up NCCT scans by one of three independent experienced stroke imaging researchers (Z.K.L., K.K., and A.A.), who were blinded to clinical data, using ITK-SNAP version 3.6.0 (http://​www.​itksnap.​org), with manual editing as required. The raters also classified each baseline NCCT scan as positive or negative for the presence of radiological markers (blend sign [7], black hole sign [8], hypodensities [9, 10], and island signs [11]). Reliability assessments for the haematoma volumetric measurement and radiological marker interpretation for these raters have been published previously [13].

Secondly, images were resampled to 1-mm isotropic voxel size and additional filtered versions were computed using Laplacian of Gaussian and Wavelet filters (see Supplementary Materials). A total of 754 NCCT radiomics-based features were subsequently extracted from the original and filtered scans (see Supplementary Table S1) using MATLAB R2019b. Seven hundred fifty-two of them were taken from the area defined as intracerebral haemorrhage on the baseline scan, using a third-party package (https://​github.​com/​mvallieres/​radiomics) [24] for textural features, and in-house code for first-order and shape-based features (see Supplementary Materials for code snippets). The remaining two features correspond to baseline perihaematomal oedema volume (mL) and baseline intraventricular haemorrhage volume (mL). Additionally, we incorporated the TICH-2 treatment allocation (either tranexamic acid or placebo) to the radiomics features as a covariate of no interest.

Each feature vector was harmonised using the MATLAB version of the ComBat harmonisation package (https://​github.​com/​Jfortin1/​ComBatHarmonizat​ion) [25‐27] with parametric adjustments to remove possible batch effects of slice thickness in the radiomics computation. This method has already been tested before for harmonisation of NCCT radiomics [28]. We utilised three batches: (1) slice thickness < 2 mm; (2) slice thickness ≥ 2 mm and < 4 mm; and (3) slice thickness ≥ 4mm. Additionally, age and gender were utilised as biological covariates. Harmonisation was performed first on the training set and the same parameters were then applied to harmonise the testing set.

Also, an iterative feature selection procedure was run in which; for every pair of variables with absolute correlation > 0.9, the one with the largest mean absolute correlation is removed. After removing each variable, the average correlations were recomputed for the next iteration. This procedure was performed on the training data only and reduced the feature set to 218 unrelated features (see Supplementary Table S2). Figure 2 depicts the feature extraction process.

We computed a generalised linear model (GLM) via elastic-net regularisation [29] with standardisation and log-loss score as energy function. To this end, we performed an exhaustive search-grid optimisation procedure with stratified 10-fold cross-validation (Fig. 3) using the H2O platform v3.26.0.2 (www.​h2o.​ai) and R software v3.6.3 (www.​r-project.​org). This search was carried out over the α blending hyperparameter of elastic-net regularisation, with values ranging from 0 (Ridge regression) to 1 (LASSO regression) in increments of 0.1 (see Supplementary Materials). Outcomes of interest were haematoma expansion, defined as volumetric growth of > 6 mL or > 33% on the follow-up scan, and poor functional outcome defined as modified Rankin scale of 4 to 6 at day 90 [20]. Finally, the optimal GLMs were chosen based on their cross-validation area under the receiver operating characteristic curve (AUC) score.

For comparison purposes, the same hyperparameter optimisation was performed using the presence of radiological signs alone as binary feature set and also using demographic and clinical information previously found to be predictive [14, 30]. The radiological markers used were the blend sign, black hole sign, hypodensities, and island signs. Demographic and clinical factors were age (years), gender (male/female), time from onset to baseline scan (hours), baseline haematoma volume (mL), antiplatelet use (yes/no), and ultra-early haematoma growth (baseline haemorrhage volume over time from onset to baseline scan). Anticoagulant use was also found to be predictive by Al-Shahi Salman et al [14]; however, this was an exclusion criterion of TICH-2 and hence not included. We also carried out the analysis using radiomics-based features combined with radiological signs and with clinical factors independently.

We also assessed our study using the radiomics quality score [31] (See Supplementary Materials for details). The score was 13 (36.11%), which is higher than the median score obtained on a recent systematic review of 51 cancer studies [32]. One of the main criteria affecting our score was the fact that despite TICH-2 being a prospective trial, it was not initially devised with radiomics in mind.

Participant demographics are summarised in Table 1 for haematoma expansion and Table 2 for poor functional outcome. Of the 1732 participants, 13 had no functional outcome information available and were excluded from the corresponding analysis. This accounts for the difference in the number of subjects between Tables 1 and 2. There was no significant difference in age, gender, or any of the included variables between training and testing sets (all p > .05). Finally, no statistical difference (all p > .05) in treatment allocation proportions was found between both sets.

Regarding differences within the training and testing sets, we observed that baseline haemorrhage volume was significantly higher (p < .001) for participants who developed HE and for those with poor functional outcome in both sets. This difference remains significant if we test on the whole population sample (p < .001). For treatment allocation, we observed a difference that is not statistically significant between participants with and without HE on both sets. However, this difference became significant when tested on the whole population sample (p = .018). We observed no statistical difference in poor functional outcome in the training and testing sets, nor when tested on the whole population sample (all p > .05).

To show the effect of feature harmonisation, we computed a 2-dimensional t-SNE manifold [33] over the standardised feature vectors of each subject for both the training and testing sets (Fig. 4). We observed that prior to harmonisation, there were three clusters corresponding to each of the harmonisation batches on both sets. This demonstrates the strong impact of slice thickness on the values of computed radiomics-based features. After harmonisation, this influence was eliminated.

Threshold analysis and ROC curves summarising the prediction performance of the optimal models on the training set using NCCT radiomics-based features, radiological signs, clinical factors, and combined models are depicted in Fig. 5 for HE and Fig. 6 for poor functional outcome. Individual AUC, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and prevalence for both the training and testing sets are reported in Table 3. For each optimal model, we chose the threshold probability in the training set such that Youden’s index defined as sensitivity + specificity − 1 is maximised. The same threshold was then used on the testing set.

For prediction of haematoma expansion, a moderate performance was observed when using NCCT radiomics-based features, with testing set AUC of 0.693 (95% CI, 0.638–0.747), sensitivity of 0.635 (95% CI, 0.554–0.716), specificity of 0.690 (95% CI, 0.644–0.736), PPV of 0.422 (95% CI, 0.355–0.49), and NPV of 0.841 (95% CI, 0.801–0.882). The rather low PPV can be partially explained by an imbalance of the positive and negative classes, as prevalence was 26.3% (95% CI, 22.5–30.1%). The most important feature in the GLM was “LoG-35 interquartile range” (Table 4), which highlights that haematoma heterogeneity (as measured by Laplacian of Gaussian-based features) is relevant for prediction. The performance of radiological signs was considerably lower (Table 3) due to sharp reductions in sensitivity and AUC on both sets; however, specificities were mildly higher. Moreover, the optimal model for prediction using radiomics-based features and radiological signs combined was identical to the optimal model using radiomics-based features alone, as all the coefficients for radiological signs were shrunk to zero by elastic-net training. Also, the optimal model using demographic and clinical factors yielded slightly lower performance than the one using radiomics-based features, especially due to a sharp reduction in sensitivity to 0.35 (95% CI, 0.27–0.43). However, combining these factors with radiomics-based features boosted their performance, achieving an AUC of 0.704 (95% CI, 0.651–0.758) and an increased sensitivity of 0.65 (95% CI, 0.57–0.73). We found that time from onset to CT scan was the factor responsible for this improvement. Finally, treatment allocation was discarded by elastic net in all radiomics-based models, showing it had no effect on predictions.

In the case of prediction of poor functional outcome using NCCT radiomics-based features, we observed a good performance, with testing set AUC of 0.783 (95% CI, 0.744–0.823), sensitivity of 0.698 (95% CI, 0.642–0.754), specificity of 0.741 (95% CI, 0.688–0.795), PPV of 0.729 (95% CI, 0.673–0.784), and NPV of 0.711 (95% CI, 0.657–0.765). Here, the balanced prevalence of 49.9% (95% CI, 45.6–54.2%) is reflected on balanced predictive values. Perihaematomal oedema volume was found as the most important factor for prediction, suggesting that inflammation may have an important effect on clinical outcome. Like the case of haematoma expansion, both sensitivity and AUC suffered a strong reduction using radiological signs compared to the model trained with NCCT radiomics-based features. Again, for the case of combined NCCT radiomics-based features and radiological signs, we obtained a model which was identical to the one trained using NCCT radiomics-based features alone. Moreover, demographic and clinical factors showed slightly better AUC than NCCT radiomics-based features, explained by an improved specificity. Again, by combining these factors and NCCT radiomics-based features, the best performance is achieved, with an AUC of 0.818 (95% CI, 0.781–0.854), mainly due to increased PPV and NPV of 0.799 (95% CI, 0.747–0.852) and 0.73 (95% CI, 0.68–0.781), respectively. The features accountable for this improvement were age and ultra-early haematoma growth. Lastly, treatment allocation had again no influence on the radiomics-based models and was discarded by elastic net.