Valuation of automatized ASPECTS as an artificial intelligence tool in daily clinical

To evaluate an automated ASPECTS (ASPECTS-a) software against two radiologists' reading of CT scans requested from the Emergency Department. Describe the most frequent failures of the ASPECTS-a. All the cranial CT Scans requested by the Emergency Department in one month were collected. The following data were recorded: age, sex, the reason for requesting the study, and imaging findings. A program was used that provides an ASPECTS score automatically. Subsequently, 2 radiologists independently reviewed all of the studies and provided the visual ASPECTS (ASPECTS-v). In case of discrepancy, a new reading was made by consensus. A total of 295 brain CT scans (45.1% male) with a mean age of 65 ± 20.0 years were included. 91.8% were interpreted as ASPECTS-v 10 in both cerebral hemispheres by both readers. ASPECTS-a scored 45% with ASPECTS 10 in both cerebral hemispheres. In 152 (51.5%) the ASPECTS-a and the ASPECTS-v did not coincide. The causes of the discrepancy were mainly due to segmentation errors (usually due to asymmetric atrophies). Most of the segmentation errors were located in the head of the caudate nucleus, observed in 60 studies. ASPECTS-a is a powerful and helpful tool, but human supervision is always necessary, particularly in groups of patients with pre-existing brain changes.