Automated, anatomy-based, heuristic post-processing reduces false positives and improves interpretability of deep...

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Deep learning (DL) models show promise for detecting intracranial aneurysms on CT angiography (CTA).
• High false positive (FP) rates hinder the clinical translation of these DL models.
• Automated methods are needed to improve the accuracy and reliability of DL-based aneurysm detection.

Purpose of the Study:

• To develop and evaluate a fully automated hybrid heuristic-DL pipeline to reduce FPs in intracranial aneurysm detection on CTA.
• To integrate automated in-scan anatomic segmentation and background/venous voxel removal into DL pipelines.
• To assess the performance of different heuristic post-processing modules in conjunction with DL models.

Main Methods:

• Two DL models, CPM-Net and a 3D-CNN-TR hybrid, were trained on 1,186 CTAs.
• A pipeline integrated DL models with heuristic post-processing modules, including brain masks and DL-based artery-vein separation.
• FPs were eliminated based on overlap with brain masks and/or vein segmentation masks.

Main Results:

• The best performing method (Method 5) reduced FPs by up to 70.6% on a private dataset and 57.9% on a public dataset, without significantly reducing true positives (TPs).
• False positive reduction lowered the false positive rate (FPR) from 0.88 to 0.26 for CPM-Net and 1.27 to 0.62 for 3D-CNN-TR on the private dataset.
• On the RSNA dataset, Method 1 effectively removed FPs while preserving TPs, demonstrating the utility of anatomy-based post-processing.

Conclusions:

• Integrating interpretable, anatomy-based modules for background and vein removal into DL pipelines improves aneurysm detection performance.
• Hybrid heuristic-DL pipelines enhance model performance and may increase clinical acceptance in radiology.
• Automated post-processing techniques are crucial for the successful clinical translation of AI in medical imaging.