AI Case Study
Qure.ai can detect critical head trauma or stroke symptoms from CT scans with more than 95% accuracy using deep neural networks and natural language processing
Qure.ai has developed algorithms that can identify bleeds, fractures and mass effect from head CT scans. Correctly diagnosing head injuries is highly time sensitive as every minute that goes by can lead to brain cell death. A convolutional neural network model was trained on over 313,000 anonymised CT scans labelled with medical diagnosis interpreted from clinical reports with natural language processing. The model can now review CT scans in under 10 seconds to detect, localise and assess the severity of injuries. The model has a 95% accuracy. And it has the potential to help reduce the effects of head trauma, improve hospital triage, and reduce costs.
Healthcare Providers And Services
Qure.ai trained "models using a collection of 313,318 anonymized head CT scans, along with their corresponding clinical reports. Of these, 21,095 scans were then used to validate the AI’s algorithms." Finally, the AI was clinically validated on 491 CT scans, with the results compared against a panel of three senior radiologists.
They used convolutional network networks to analyse the CT scans. They also used natural language processing (NLP) to analyse clinical radiology reports to label the scans with appropriate head injury information.
"NLP and machine vision solution can automatically generate abnormality reports for CT scans. CT scan technology rapidly screens scans in under 10 seconds to detect, localize and quantify abnormalities, as well as assess their severity. These reports help radiologists and hospitals prioritize care, make smarter and faster diagnoses and reduce costs which is critical especially for stroke patients, as every minute that goes by brain cells die."
R And D
Core Research And Development
"...head CT scans are among the most commonly used emergency room diagnostic tools for patients with head injury or in those with symptoms suggesting a stroke or rise in intracranial pressure. Their wide availability and relatively low acquisition time makes them a commonly used first-line diagnostic modality.
Detecting 'the most critical, time-sensitive abnormalities that can be readily detected on CT scan include intracranial hemorrhages, raised intracranial pressure and cranial fractures. A key evaluation goal in patients with stroke is excluding an intracranial hemorrhage. This depends on CT imaging and its swift interpretation. Similarly, immediate CT scan interpretation is crucial in patients with a suspected acute intracranial hemorrhage to evaluate the need for neurosurgical treatment. Cranial fractures, if open or depressed will usually require urgent neurosurgical intervention. Cranial fractures are also the most commonly missed major abnormality on head CT scans especially if coursing in an axial plane.
While these abnormalities are found only on a small fraction of CT scans, streamlining the head CT scan interpretation workflow by automating the initial screening and triage process, would significantly decrease the time to diagnosis and expedite treatment. This would in turn decrease morbidity and mortality consequent to stroke and head injury. An automated head CT scan screening and triage system would be valuable for queue management in a busy trauma care setting, or to facilitate decision-making in remote locations without an immediate radiologist availability."
"The percentage of annual US emergency room visits that involve a CT scan has been increasing for the last few decades and the use of head CT to exclude the need for neurosurgical intervention is on the rise."
"The validation study found that Qure.ai’s AI was more than 95% accurate in identifying abnormalities."
"The AI was clinically validated on 491 CT scans, with the results compared against a panel of three senior radiologists."
"Qure.ai released a clinical validation study showing its algorithms were nearly on par with radiologists in a sample of 21,000 patients.
"further research is necessary to determine if these algorithms enhance the radiologists’ efficiency and ultimately improve patient care and outcome."
Convolutional neural networks were used to analyse the CT scans. "Training this model requires large amount of data for which the ground truth is already known...One of the main challenges we faced in the development of the algorithms was the three dimensional (3D) nature of the CT scans. This was primarily due to an issue termed as ‘curse of dimensionality’ where the data required to train a machine learning algorithm scales exponentially with the dimensionality of data."
"A natural language processing (NLP) algorithm was used to detect IPH, SDH, EDH, SAH, IVH, calvarial fractures from clinical radiology reports."
"Qure.ai trained the new AI using a collection of 313,318 anonymized head CT scans, along with their corresponding clinical reports. Of these, 21,095 scans were then used to validate the AI’s algorithms. Finally, the AI was clinically validated on 491 CT scans [with 193,318 slices], with the results compared against a panel of three senior radiologists."
Of these scans, we earmarked scans of 23,163 randomly selected patients (Qure25k dataset) for validation and used the scans of rest of the patients (development dataset) to train/develop the
algorithms. We removed post-operative scans and scans of patients less than 7 years old from the Qure25k dataset. This dataset was not used during the algorithm development process."