Метадеректерді қарау

BACKGROUND: Chest X-ray examination is one of the first radiology areas that started applying artificial intelligence, and it is still used to the present. However, when interpreting X-ray scans using artificial intelligence, radiologists still experience several routine restrictions that should be considered in issuing a medical report and require the attention of artificial intelligence developers to further improve the algorithms and increase their efficiency.

AIM: To identify restrictions of artificial intelligence services for analyzing chest X-ray images and assesses the clinical significance of these restrictions.

MATERIALS AND METHODS: A retrospective analysis was performed for 155 cases of discrepancies between the conclusions of artificial intelligence services and medical reports when analyzing chest X-ray images. All cases included in the study were obtained from the Unified Radiological Information Service of the Unified Medical Information and Analytical System of Moscow.

RESULTS: Of the 155 analyzed difference cases, 48 (31.0%) were false-positive and 78 (50.3%) were false-negative cases. The remaining 29 (18.7%) cases were removed from further studies because they were true positive (27) or true negative (2) in the expert review. Most (93.8%) of the 48 false-positive cases were due to the artificial intelligence service mistaking normal chest anatomy (97.8% of cases) or catheter shadow (2.2% of cases) for pneumothorax signs. Overlooked clinically significant pathologies accounted for 22.0% of false-negative scans. Nearly half of these cases (44.4%) were overlooked lung nodules. Lung calcifications (60.9%) were the most common clinically insignificant pathology.

CONCLUSIONS: Artificial intelligence services demonstrate a tendency toward over diagnosis. All false-positive cases were associated with erroneous detection of clinically significant pathology: pneumothorax, lung nodules, and pulmonary consolidation. Among false-negative cases, the rate of overlooked clinically significant pathology was low, which accounted for less than one-fourth.