Machine Learning And Visual Recognition In Computer-Aided Diagnosis: A Proof-of-Concept Acute Intracranial Hemorrhage Screening And Classification Model
Alejandra M Casar Berazaluce, Brandon E Colvin, Todd A Ponsky, *Richard A Falcone, Jr.
Cincinnati Children's Hospital Medical Center, Cincinnati, OH
BACKGROUND Machine learning enables the use of photographs, microscopy, and radiological studies for computer-aided diagnosis. With increasing use of imaging, formal reads may take time to obtain or be unavailable after hours. During acute resuscitations, findings may be missed due to inexperience or clinical distractors. Delays in recognition of intracranial hemorrhage can significantly impact patient management and outcomes. This study evaluates the feasibility of creating diagnostic models using IBM Watson.
METHODS Axial cuts of non-contrasted head CTs were selected for analysis. A Google Image search was performed for normal head CT and subdural, epidural, and intraparenchymal hemorrhage. 180 images were collected; 160 were used for training and 20 for testing. Two machine learning algorithms were developed.
RESULTS Our screening model was 100% accurate in differentiating normal and abnormal scans. The classification model had mean sensitivity of 90%, specificity of 92%, PPV of 81%, and NPV of 97% for identifying individual diagnoses; with 100% accuracy for normal, reinforcing the success of the screening model.
CONCLUSIONS Custom visual recognition models using machine learning are a feasible pathway to creating screening or diagnostic tools. Being able to automatically detect abnormal findings represents an area of relatively untapped potential for improving patient safety and quality of care. Further studies using higher quality data with a variety of diagnoses and imaging modalities are warranted to explore the possibilities afforded by this technology.
Back to 2019 Abstracts