Published in the European Medical Journal and based on 372 patients, research demonstrated that automated computed tomography radiomics holds potential for forecasting unfavourable functional recovery following primary brainstem haemorrhage.
While the combined model outperformed conventional radiological assessment for functional outcomes, its ability to predict post-stroke pneumonia was less consistent.
Researchers developed an automated segmentation-based radiomics framework using emergency non-contrast computed tomography scans. The retrospective study included 372 patients with primary brainstem haemorrhage, divided into training, internal testing, and temporal internal validation datasets. To isolate the haematomas, a three-dimensional U-Net was employed, allowing investigators to evaluate radiomics models derived from both automated and manual segmentation techniques before merging automated radiomics attributes with standard radiological predictors.
Across all three cohorts, the radiomics model utilizing automated segmentation yielded results comparable to the manual segmentation approach. Regarding the forecast of poor functional recovery, the integrated model attained area under the curve metrics of 0.926 for the training set, 0.923 for the internal test set, and 0.858 for the temporal internal validation set.
These results represented higher area under the curve values and lower Brier scores than the conventional radiological model, indicating improved predictive performance. Using SHapley Additive exPlanations, researchers identified haematoma classification as the most influential predictor in the prognostic model.
Conversely, the combined post-stroke pneumonia model yielded area under the curve metrics of 0.828, 0.742, and 0.681 throughout the respective three datasets, showing no steady performance enhancement. The findings suggest that automated segmentation-based non-contrast computed tomography radiomics could support early risk stratification in patients, though the post-stroke pneumonia prediction model requires further refinement and validation.
Researchers concluded that automated computed tomography radiomics is feasible for early prognostic assessment, with further work needed to establish its predictive value.