Automated triage of acute stroke cases: Integrating imaging and clinical information to prioritize clinical interpretation
Stroke is the third leading cause of death in the United States. The timely detection and treatment of stroke is important in reducing disability and death. The diagnosis of acute stroke in the clinical setting involves an initial neurological examination. However, the clinical presentation of other neurological disorders may mimic stroke, making it difficult to establish a diagnosis of stroke based on clinical findings alone. Neuroimaging has been identified as a key component to the initial evaluation of acute stroke since it can confirm the presence or absence of acute stroke. The inclusion of neuroimaging in the initial evaluation requires the radiologist to play a prominent role in the triaging of stroke patients.
In order for radiologists to provide a final clinical diagnosis of acute stroke, several different sets of images and clinical information must be reviewed and analyzed. With advances in imaging modalities, sequences, and technology, the number of images available for triaging stroke patients has increased tremendously. The radiologist is confronted with the challenge of organizing and analyzing a large volume of images in order to provide timely diagnosis. Current image storage and management systems such as Picture Archival and Communications Systems (PACS) generally provide radiologists with a first in, first out (FIFO) or a last in, first out (LIFO) method for organizing imaging cases rather than ranking images based on urgency. This requires the radiologist to manually review all image datasets acquired, including non-stroke cases, in order to detect an acute stroke. This can potentially increase the time between image acquisition and the final stroke diagnosis, which can lead to a delay in the treatment and management of stroke patients.
The objective of this thesis is to develop automated methods that can be integrated into a fully automated platform for the triaging of radiological images. Since radiologists utilize images and clinical information in diagnosing acute infarcts, a set of automated techniques using pattern recognition, natural language processing, and machine learning is implemented and applied to both imaging and clinical information.
The results from these automated techniques indicate that triaging acute infarct cases based on combined imaging and clinical data outperforms triaging based on clinical data alone, imaging data alone or random triaging. The automated methods developed in this thesis can assist the radiologist in triaging a large number of cases and can potentially reduce the time between image acquisition, interpretation, and final diagnosis.