A deep-learning algorithm can provide an effective second assessment of MR angiography (MRA) studies, helping to catch cerebral aneurysms that were missed by the radiologist during initial interpretation, Japanese researchers reported in an article published online October 22 in Radiology.
The research group led by Daijo Ueda of Osaka City University Graduate School of Medicine trained a deep-learning algorithm to automatically detect cerebral aneurysms on time-of-flight (TOF) MRA studies. In testing, the algorithm yielded 93% sensitivity compared with the initial radiology reports. Importantly, though, it found up to 13% more aneurysms than had been detected by the original radiologist interpretations.
"By providing a second assessment of the images, our algorithm could not only help radiologists to detect cerebral aneurysms, but also reduce the risk of overlooking aneurysms," the authors wrote.
In 85% of cases, nontraumatic subarachnoid hemorrhages are caused by aneurysms. TOF MRA has been reported to have 87% sensitivity for cerebral aneurysms. To provide a second assessment of images already interpreted by radiologists, the researchers sought to develop a sensitive and versatile deep-learning algorithm by using training cases produced at multiple institutions by different scanner types.
"A highly versatile algorithm such as this is desirable because the appearance of aneurysms differs depending on the type of imaging device used and the conditions under which the TOF MR angiography was performed," the authors wrote. "Additionally, there is variability in the diagnoses between individual radiologists."
They trained a ResNet-18 convolutional neural network using a dataset of 748 aneurysms from 683 MRA examinations at two clinics and two hospitals using three different types of MRI scanners from two different vendors. All images had been independently interpreted by at least two radiologists during clinical routine; in the event of discrepant diagnoses, the final decision was made in discussion between the two radiologists. The researchers then tested the performance of the algorithm on two different datasets: an internal testing dataset of 649 aneurysms collected from their hospital and an external dataset of 80 aneurysms from another clinic.
| Algorithm's performance compared with initial radiologist interpretations | ||
| Internal testing dataset | External dataset | |
| Sensitivity | 592/649 (91%) | 74/80 (93%) |
| Aneurysms detected by algorithm but missed by radiologists | 31/649 (4.8%) | 10/80 (13%) |
The researchers noted that aneurysms missed by the algorithm tended to be larger (at least 10 mm) and in the vertebral artery; the relatively low sensitivities may have been caused by a lack of these cases in the training dataset. Furthermore, the internal signal in these missed lesions was more heterogeneous than in the other aneurysms.
"Because aneurysms with a homogeneous internal signal were correctly detected by the algorithm, regardless of size and location, we thought that heterogeneity in the flow of the aneurysm also decreased detectability," the authors wrote. "These heterogeneous aneurysms are usually not missed by radiologists because of their size and irregularity. This may suggest that radiologists and the present algorithm with deep-learning technique are complementary to each other."
