Noncontrast CT-based habitat analysis for differentiating hemorrhagic transformation from contrast extravasation after endovascular therapy of acute ischemic stroke.
Huang Y, Tan G, et al. • European radiology experimental • 2026
Habitat analysis outperformed radiomics analysis in differentiating hemorrhagic transformation from contrast extravasation within hyperdense areas after endovascular therapy on noncontrast CT, achieving AUCs of 0.89 and 0.82 in training and validation cohorts respectively.
Key Findings
Results
The habitat model achieved the highest diagnostic performance for differentiating HT from CEx, with AUCs of 0.89 and 0.82 in the training and validation cohorts respectively.
The study included 187 AIS patients with hyperdense areas (HDAs) after EVT in a retrospective design
A stratified three-fold cross-validation was used to split the dataset into training and validation cohorts
Habitat analysis was applied to postoperative noncontrast CT images
The habitat model AUC of 0.89 in training and 0.82 in validation cohorts represented the top-performing model tested
Results
Habitat analysis demonstrated improved diagnostic performance compared with radiomics analysis for differentiating HT from CEx.
Radiomics model achieved AUCs of 0.81 and 0.75 in the training and validation cohorts, respectively
Habitat model AUCs were 0.89 (training) and 0.82 (validation), representing improvements of 0.08 and 0.07 over radiomics
Both models were applied to postoperative noncontrast CT images
Habitat analysis outperformed radiomics analysis in both training and validation cohorts
Methods
The optimal number of clusters (k-value) for habitat analysis was determined to be 5 using the k-means clustering algorithm.
The k-means clustering algorithm was applied to postoperative noncontrast CT
The optimal k-value was determined to be 5
Five-cluster habitat analysis revealed intralesional heterogeneity within HDAs in AIS patients
This clustering approach enabled characterization of spatial heterogeneity within the hyperdense areas
Results
The combined model integrating preoperative imaging features, clinical characteristics, and habitat features did not improve diagnostic performance over the habitat model alone.
Preoperative hyperdense artery sign (HAS) status and clinical characteristics were collected and incorporated into combined models
Combined models integrating HAS and clinical characteristics with habitat features were developed and evaluated
The addition of preoperative imaging features and clinical characteristics did not improve upon the habitat model's performance
Both radiomics and habitat models had corresponding combined model variants tested
Results
Five-cluster habitat analysis revealed intralesional heterogeneity within hyperdense areas in AIS patients after EVT.
Habitat analysis was performed on noncontrast CT, which is widely available in most stroke centers
The analysis revealed spatial heterogeneity within the HDAs that single-region radiomics analysis does not capture
The study authors describe this as supporting habitat analysis's 'potential as a novel tool for clinical decision-making'
Differentiation of HT from CEx in post-EVT HDA on noncontrast CT is described as 'a clinical challenge in AIS'
What This Means
After a stroke is treated with a procedure to remove or dissolve the blood clot (called endovascular therapy), doctors sometimes see bright white areas on CT scans of the brain. These bright areas can represent either a dangerous bleeding complication called hemorrhagic transformation, or a harmless remnant of the contrast dye used during the procedure called contrast extravasation. Telling these two apart is critical because they require very different management, but it is difficult to do with standard CT imaging alone. This study tested a new computer-based method called 'habitat analysis' that divides these bright areas into five distinct zones based on their image characteristics, effectively mapping the internal complexity of the lesion.
The researchers found that habitat analysis was better than standard radiomics (another computer-based image analysis method) at correctly identifying whether a bright area was bleeding or leftover dye. The habitat approach achieved accuracy scores (AUC) of 0.89 in the development group and 0.82 in the testing group, compared to 0.81 and 0.75 for standard radiomics. Interestingly, adding information from pre-procedure scans and patient clinical details did not further improve accuracy beyond what habitat analysis alone could provide. The analysis was performed on standard noncontrast CT scans, which are routinely available in virtually all hospitals that treat stroke.
This research suggests that habitat analysis could be a practical new tool to help clinicians quickly and accurately determine what is causing bright spots on post-treatment brain CT scans, without requiring additional imaging or contrast agents. Because it works on the standard CT scans already obtained after stroke treatment, this approach could potentially be implemented in most stroke centers without major additional resources. The findings support further investigation of habitat analysis as a decision-support tool in acute stroke care.
Huang Y, Tan G, Fan W, Yang L, Zhang T, Zhang Y, et al.. (2026). Noncontrast CT-based habitat analysis for differentiating hemorrhagic transformation from contrast extravasation after endovascular therapy of acute ischemic stroke.. European radiology experimental. https://doi.org/10.1186/s41747-026-00798-9