बायोमेडिकल इंजीनियरिंग और मेडिकल डिवाइसेस जर्नल
खुला एक्सेस

अमूर्त

Artefact-tolerant Intracranial Haemorrhage Segmentation Method of Non-contrast CTs: An Open-sourced Tool and Dataset for Algorithm Development

Antonios Konstantinos Thanellas, Mikko Lilja, Nik Lygeros, Teijo Kottila, Miikka Korja

Objectives: We aimed to create an artefact-tolerant and fully automated segmentation method intended to reduce the
workload of medical experts who segment head computed tomography images of intracranial haemorrhage patients.
Methods: We developed a segmentation algorithm that combines 2D and 3D intensity thresholding, morphological
operations, and entropy filtering. We tested the algorithm’s performance against gold standard segmentations on
preoperative and postoperative/posttreatment head computed tomography images of 145 patients with intracranial
bleeding. We compared the fully automated algorithm against a simpler thresholded method.
Results: The fully automated algorithm correctly segmented blood in 98.62% of patients, in 2277 out of 2449
positive slices (92.97%), and in 54.12% of positive voxels. It incorrectly segmented blood in 0.63% of patients’
negative voxels. The Dice coefficient at voxel level was 0.20.
Conclusion: The open-sourced algorithm may facilitate the segmentation of a wide quality range of preoperative or
postoperative/posttreatment head computed tomography scans with intracranial haemorrhage.