Automatic intracranial abnormality detection and localization in head CT scans by learning from free-text reports

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Deep learning models for medical image diagnosis require extensive manual annotations, which are costly and time-consuming.
• Current methods face challenges in efficiently utilizing the rich information present in free-text clinical reports.

Purpose of the Study:

• To develop a novel cross-modality learning framework (Cross-DL) for intracranial abnormality detection and localization in head CT scans.
• To leverage free-text imaging reports to overcome the limitations of manual image annotation.

Main Methods:

• Cross-DL employs a discretizer to automatically extract abnormality labels (types and locations) from free-text reports.
• An image analyzer is trained using a dynamic multi-instance learning approach with these extracted labels.
• The framework utilizes a large-scale dataset of 28,472 head CT scans.

Main Results:

• Cross-DL achieved an average area under the receiver operating characteristic curve (AUROC) of 0.956 for detecting 4 abnormality types in 17 regions.
• The model demonstrated accurate voxel-level localization of abnormalities.
• An external validation on the CQ500 dataset for intracranial hemorrhage classification yielded an AUROC of 0.928.

Conclusions:

• Cross-DL effectively utilizes free-text reports for training deep learning models in medical imaging, significantly reducing annotation costs.
• The framework demonstrates high performance in detecting and localizing intracranial abnormalities on CT scans.
• This approach offers a scalable solution for medical image analysis and can aid in prioritizing radiological reviews.