Deep Learning Models for Automatic Classification of Anatomic Location in Abdominopelvic Digital Subtraction...

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Vascular Interventional Radiology

Background:

• Digital subtraction angiography (DSA) is crucial for visualizing vascular anatomy.
• Identifying specific anatomic locations within DSA sequences can be challenging due to information sparsity.
• Automated methods are needed to enhance efficiency and accuracy in DSA interpretation.

Purpose of the Study:

• To evaluate the information content of routine DSA.
• To assess deep learning algorithms for automated anatomic identification in DSA.
• To improve the management of information sparsity in vascular imaging.

Main Methods:

• Retrospective analysis of DSA sequences from endovascular procedures (2010-2020).
• Development of Mode and Multiple Instance Learning (MIL) models for anatomic location classification.
• Performance evaluation based on multiclass classification accuracy using key images and full datasets.

Main Results:

• Mode and MIL deep learning models achieved high accuracy (0.975 and 0.966, respectively).
• Both models significantly outperformed a baseline Mode model trained on the full dataset.
• The MIL model automatically identified key images with high overlap to manually labeled ones.

Conclusions:

• Deep learning algorithms demonstrate high fidelity in identifying abdominopelvic DSA anatomic locations.
• Manual or automatic methods can effectively manage information sparsity in DSA.
• AI-powered analysis holds significant promise for improving DSA interpretation.