Automated ASPECTS Segmentation and Scoring Tool: a Method Tailored for a Colombian Telestroke Network

Area of Science:

• Radiology
• Medical Imaging
• Neurology

Background:

• Early detection of ischemic infarcts is crucial for acute ischemic stroke management.
• Telestroke software can improve access to stroke expertise, but requires reliable diagnostic tools.
• Existing methods for infarct detection may have limitations in accuracy or accessibility.

Purpose of the Study:

• To evaluate two novel non-machine learning (non-ML) algorithmic approaches for early ischemic infarct detection on brain CT images.
• To assess the performance of these methods against a commercial ML solution and expert neuroradiologists.
• To determine the suitability of these algorithms for integration into telestroke software, particularly in resource-limited settings.

Main Methods:

• Developed and tested two non-ML algorithms: Mean Hounsfield Unit (HU) relative difference (RELDIF) and density distribution equivalence test (DDET).
• Utilized CT images from 113 acute stroke patients, with expert consensus as the gold standard for Alberta Stroke Program Early CT Scores (ASPECTS).
• Compared RELDIF and DDET performance against a commercial ML solution (CMLS) and four independent neuroradiologists using dichotomized-ASPECTS.

Main Results:

• For dichotomized-ASPECTS (<6), DDET achieved an area under the receiver operating characteristic curve (AUC) of 0.85, and RELDIF achieved 0.84, outperforming the CMLS (0.64).
• Accuracy for DDET was 0.85 and for RELDIF was 0.88, comparable to neuroradiologists (0.83-0.96).
• Both DDET and RELDIF demonstrated equivalence to the gold standard and non-inferiority compared to CMLS and at least one neuroradiologist.

Conclusions:

• The RELDIF and DDET methods show high accuracy and reliability in detecting early ischemic infarcts on brain CT.
• These non-ML algorithms hold significant potential as supportive tools for prompt and accurate stroke diagnosis in clinical settings.
• Their performance suggests they could be valuable additions to telestroke software, especially in regions with limited access to neuroradiologists.