Accurate and rapid quantification of intracerebral haemorrhage (ICH), perihematomal oedema (PHE) and intraventricular haemorrhage (IVH) plays a pivotal role in determining treatment strategies and predicting patient outcomes. Manual methods such as the ABC/2 technique have long been used, but they lack precision and consistency, especially for complex or small-volume lesions. This limitation can lead to misjudged interventions or treatment delays. Addressing these gaps, a multicentre study developed a nnU-Net-based deep learning model capable of automated multilabel segmentation on non-contrast CT (NCCT). This model demonstrates strong potential to transform clinical workflows through accurate, fast and reproducible assessments of acute haemorrhagic lesions.
Model Design and Validation Across Multicentre Datasets
The model was trained using data from 775 patients and tested on 189 independent cases, supplemented by internal (n=121) and external (n=169) validation sets. It was developed using the nnU-Net framework, known for its self-configuring capabilities in biomedical image segmentation. Training followed a 5-fold cross-validation approach with hyperparameters tailored for medical image segmentation. The model was implemented on high-performance GPUs and assessed against manual segmentations that served as the reference standard.
Recommended Read: Enhancing Brain MRI Diagnosis: The Role of Human-AI Collaboration
Validation across independent cohorts confirmed robust generalisability. Internal validation yielded Dice Similarity Coefficients (DSC) of 0.88 for ICH, 0.66 for PHE and 0.80 for IVH, while external datasets showed similar trends. Processing times were drastically reduced, with automated segmentation averaging just 18.2 seconds compared to over 18 minutes for manual annotations. The model maintained high sensitivity and specificity across lesion types, with detection rates exceeding 97% for ICH and PHE and performing strongly for IVH, even at volumes as low as 0.2 mL.
Performance Across Lesion Types and Anatomical Regions
Segmentation accuracy varied slightly across lesion types and anatomical locations but consistently demonstrated clinical relevance. For ICH, the median DSC was 0.91 in the test set, reflecting strong agreement with manual ground truth. PHE, known for its diffuse and less well-defined nature, achieved a median DSC of 0.71. While slightly lower, this score still reflected a high correlation, supported by a Pearson coefficient of 0.92. IVH segmentation, more susceptible to anatomical overlap and morphological variability, achieved a median DSC of 0.76 for volumes over 1 mL, improving upon general metrics that included sub-millimetre lesions.
Location-specific analyses indicated strong ICH performance in lobar and deep regions, with DSC values of 0.90 and 0.92, respectively. Performance in the brainstem and cerebellum declined due to case rarity and structural complexity, suggesting a need for targeted training data augmentation. Nonetheless, volumetric estimates across all lesion types correlated strongly with manual measurements, reinforcing the model’s reliability for clinical decision-making.
Clinical Applicability and Implementation Challenges
The nnU-Net model represents a major step toward automating volumetric analysis in acute stroke care. By delivering precise measurements quickly, it enables timely therapeutic interventions, such as blood pressure management, surgical decision-making or anticoagulation reversal. Furthermore, automated tracking of lesion evolution facilitates monitoring treatment efficacy and outcome prognostication over time.
However, real-world implementation requires addressing practical constraints. High-performance computing infrastructure, integration with radiology systems and radiologist training are necessary for routine use. Ensuring DICOM compatibility, managing data security and calibrating outputs to site-specific imaging protocols are also essential for widespread deployment. Despite these hurdles, the model’s strong cross-site validation and reproducibility highlight its readiness for clinical translation, especially in high-volume stroke centres.
The study presents a robust and generalisable deep learning model for automated segmentation of ICH, PHE and IVH on NCCT. Trained and validated across multiple institutions and scanner types, the model demonstrates high accuracy and efficiency, offering significant advantages over manual approaches. Its capacity to deliver fast, standardised and volumetrically accurate assessments holds great promise for improving clinical workflows, treatment planning and outcome prediction in patients with spontaneous ICH. Continued refinement, prospective validation and integration into clinical systems will be key to realising its full impact in acute neuroimaging.
Source: Radiology Advances
Image Credit: Freepik
