Assistance from Automated ASPECTS Software Improves Reader Performance

Area of Science:

• Neurology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate Alberta Stroke Program Early CT Score (ASPECTS) assessment is crucial for large vessel occlusion (LVO) stroke management.
• Early CT interpretation can be challenging, especially for non-neuroradiologists.

Purpose of the Study:

• To evaluate the impact of a machine learning (ML)-based software tool on physicians' ability to interpret ASPECTS.
• To compare unassisted ASPECTS scoring with ML-assisted scoring in LVO stroke patients.

Main Methods:

• 50 CT scans from prior studies were assessed by 3 expert neuroradiologists to establish a reference standard.
• Non-neuroradiologist readers and neuroradiologists scored scans with and without ML software assistance.
• Agreement was measured by the percentage of ASPECTS regions correctly identified.

Main Results:

• Unassisted typical readers achieved 72% agreement with the expert consensus.
• The ML software alone achieved 77% agreement.
• ML-assisted reading improved agreement to 78%, a statistically significant enhancement (P<0.0001).

Conclusions:

• Machine learning software significantly improves ASPECTS scoring accuracy for typical readers.
• ML-assisted interpretation allows non-expert readers to achieve agreement rates comparable to neuroradiologists.