Design and optimization of an automatic deep learning-based cerebral reperfusion scoring (TICI) using thrombus...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• The Thrombolysis in Cerebral Infarction (TICI) scale is crucial for assessing mechanical thrombectomy outcomes.
• Significant variability exists in TICI scoring, necessitating improved assessment methods.

Purpose of the Study:

• To develop and optimize an artificial intelligence (AI)-based classification model for digital subtraction angiography (DSA) TICI scoring.
• To evaluate the AI model's performance in classifying TICI scores into two and three distinct groups.

Main Methods:

• A convolutional neural network (CNN) was trained on a monocentric DSA dataset of thrombectomies.
• The model classified TICI scores into two groups (TICI 0-2a vs. 2b-3) and three groups (TICI 0-2a vs. 2b vs. 2c-3).
• Thrombus positions were introduced manually and via an automated detection module to assess their impact on performance.

Main Results:

• The AI model achieved high specificity (0.97 ± 0.01) and sensitivity (0.86 ± 0.01) for the two-class TICI classification.
• The three-class classification model demonstrated insufficient performance, with F1 scores for TICI 2b around 0.50-0.55.
• Automatic thrombus detection did not significantly improve the performance of the three-class model.

Conclusions:

• The AI model offers a reproducible method for two-class TICI scoring in DSA.
• The current AI model's performance is inadequate for distinguishing three TICI grades in clinical practice.
• Automated thrombus detection did not enhance the AI model's accuracy for multi-class TICI assessment.