Deep learning for prediction of post-thrombectomy outcomes based on admission CT angiography in large vessel...

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Stroke Neurology

Background:

• Computed Tomography Angiography (CTA) is crucial for diagnosing Large Vessel Occlusion (LVO) strokes.
• Predicting 3-month outcomes after LVO thrombectomy is essential for patient management.
• Current outcome prediction methods may be limited in certain clinical situations.

Purpose of the Study:

• To develop and validate automated deep learning models for predicting 3-month outcomes in anterior circulation LVO stroke patients.
• To assess the performance of models using different input data combinations: CTA alone, CTA with treatment information, and CTA with treatment and clinical data.

Main Methods:

• Trained end-to-end deep learning pipelines on a dataset of 591 patients, split into training/validation and independent test sets.
• Utilized a pre-trained ResNet-50 3D Convolutional Neural Network (MedicalNet) with CTA preprocessing.
• Compared models incorporating "CTA" images, "CTA + Treatment" (thrombectomy time, reperfusion success), and "CTA + Treatment + Clinical" (age, sex, NIHSS).

Main Results:

• The ensemble model achieved an Area Under the Curve (AUC) of 0.70 for CTA alone.
• Models incorporating treatment and clinical data showed improved performance, with AUCs of 0.79 and 0.86, respectively.
• A logistic regression model with "Treatment + Clinical" data also achieved a high AUC of 0.86.

Conclusions:

• Demonstrated the feasibility of automated deep learning models for predicting stroke outcomes from admission CTA and post-thrombectomy data.
• These models can significantly aid prognostication in telehealth transfers and situations with communication barriers.
• The integration of clinical and treatment data enhances predictive accuracy for patient outcomes.