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

Abstract

Objective: To create a rapid, cloud-based, and deployable machine learning (ML) method named hemorrhage evaluation and detector system for underserved populations, potentially across the Mayo Clinic enterprise, then expand to involve underserved areas and detect the 5 subtypes of intracranial hemorrhage (IH). Methods: We used Radiological Society of North America dataset for IH detection. We made 4 total iterations using Google Cloud Vertex AutoML. We trained an AutoML model with 2000 images, followed by 6000 images from both IH positive and negative classes. Pixel values were measured by the Hounsfield units, presenting a width of 80 Hounsfield and a level of 40 Hounsfield as the bone window. This was followed by a more detailed image preprocessing approach by combining the pixel values from each of the brain, subdural, and soft tissue window-based gray-scale images into R(red)-channel, G(green)-channel, and B(blue)-channel images to boost the binary IH classification performance. Four experiments with AutoML were applied to study the effects of training sample size and image preprocessing on model performance. Results: Out of the 4 AutoML experiments, the best-performing model was the fourth experiment, where 95.80% average precision, 91.40% precision, and 91.40% recall were achieved. On the basis of this analysis, our binary IH classifier hemorrhage evaluation and detector system for underserved populations appeared both accurate and performed well. Conclusion: Hemorrhage evaluation and detector system for underserved populations is a rapid, cloud-based, deployable ML method to detect IH. This tool can help expedite the care of patients with IH in resource-limited hospitals.