Deep learning with weak annotation from diagnosis reports for detection of multiple head disorders: a prospective,...

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Deep learning for medical diagnostics

Background:

• High-quality annotated datasets are crucial but expensive for medical AI.
• Existing methods struggle with the cost and complexity of expert annotations.
• This work addresses the need for accurate AI without manual data labeling.

Purpose of the Study:

• To develop a deep learning system for detecting multiple head disorders from CT scans.
• To achieve accurate and generalizable performance using only weak annotations from diagnosis reports.
• To overcome the limitations of traditional annotation methods in medical imaging.

Main Methods:

• Developed a deep learning system using 104,597 head CT scans and diagnosis reports.
• Implemented a novel weakly supervised learning algorithm (RoLo) with noise tolerance and multi-instance learning.
• Validated the system on diverse retrospective, prospective, cross-center, cross-equipment, and cross-nation datasets.

Main Results:

• Achieved high area under the ROC curve (0.964-0.976) across various test datasets.
• Demonstrated performance comparable to four radiologists.
• Improved radiologist sensitivity by 0.109 and specificity by 0.022 when aided by the system.

Conclusions:

• The developed system achieves accurate and generalizable head disorder detection without expert-annotated data.
• The system enhances radiologist diagnostic performance, improving accuracy and efficiency.
• This computer-aided diagnostic system shows potential for clinical application in diverse healthcare settings.