Abstract Body: Background and Aims:
Machine learning models have shown promising potential for automated hemorrhage detection and segmentation, alleviating highly time-consuming manual contouring and facilitating rapid clinical diagnosis. External validation is essential to assess model generalizability and performance across new dataset configurations. To this end, we externally validated a novel model for hemorrhage detection and segmentation in an unseen randomized-controlled trial dataset.
Methods: A novel segmentation architecture based on denoising diffusion probabilistic models was utilized for segmenting hemorrhage. The model had been trained upon a set of 331 CT scans with manually segmented parenchymal hemorrhage lesions. External validation was conducted using the AcT (Alteplase compared to Tenecteplase) trial in which patients underwent post-thrombolysis follow-up CT scans. Ground truth regarding hemorrhage presence was determined by expert readers blinded to the algorithm’s results who performed manual contouring and graded hemorrhages using the Heidelberg classification. Model performance was then evaluated through using diagnostic performance measures and the Dice coefficient.
Results: Among the 1338 patients with follow-up CT scans, two types of hemorrhages were adjudicated: (a) any kind of hemorrhage (230/1338) and (b) remote or local hemorrhages classified as PH1 or worse (170/1338). The algorithm achieved sensitivity of 89.4% (95% CI 84.8-94.1%) for hemorrhages ≥PH1 and 61.7% (95% CI 55.5-68.0%) for any kind of hemorrhage where specificity of 92.8% (95% CI 91.3-94.3%), positive predictive value of 64.0% (95% CI 57.6-70.3%), negative predictive value of 92.1% (95% CI 90.5-93.7%) and accuracy of 87.4% was achieved. Dice was 0.582 (95% CI 0.538, 0.629) for any kind of hemorrhage and 0.611 (0.555, 0.667) for hemorrhages ≥PH1.
Conclusions: Our automated model for hemorrhage segmentation demonstrated robust performance in this external clinical trial dataset, achieving high sensitivity for large hemorrhages and high specificity for smaller hemorrhages. Future work will seek to further optimize the algorithm’s performance for detection and segmentation of smaller hemorrhages.