Hemorrhage Evaluation and Detector System for Underserved Populations: HEADS-UP

Area of Science:

• Medical Imaging
• Machine Learning
• Radiology

Background:

• Intracranial hemorrhage (IH) detection is critical for patient outcomes.
• Existing diagnostic tools may be inaccessible in underserved or resource-limited settings.
• There is a need for rapid, deployable methods for IH detection.

Purpose of the Study:

• To develop a rapid, cloud-based machine learning (ML) method for detecting intracranial hemorrhage (IH).
• To create a deployable system, the hemorrhage evaluation and detector system for underserved populations, for potential use across healthcare enterprises and underserved areas.
• To enable detection of the 5 subtypes of IH.

Main Methods:

• Utilized the Radiological Society of North America dataset for IH detection.
• Employed Google Cloud Vertex AutoML, conducting four iterative experiments with increasing dataset sizes (2000 to 6000 images).
• Implemented advanced image preprocessing by combining windowed grayscale images into RGB channels to enhance IH classification.

Main Results:

• The best-performing model achieved 95.80% average precision, 91.40% precision, and 91.40% recall.
• Analysis demonstrated the accuracy and effectiveness of the binary IH classifier.
• The study confirmed the impact of training sample size and image preprocessing on model performance.

Conclusions:

• The hemorrhage evaluation and detector system for underserved populations is a validated, rapid, cloud-based ML tool for IH detection.
• This system can significantly expedite care for patients with IH, particularly in resource-limited hospitals.
• The developed tool shows promise for improving diagnostic capabilities in underserved populations.