XXX. Magyar Radiológus Kongresszus

3. Szekció: Mesterséges intelligencia

S03.01: Evaluation of an AI based software solution in identifying normal chest X-ray studies.

Naglis Ramanauskas, Darius Barušauskas, Oxipit Ltd, Vilnius, Lithuania

Összefoglaló szövege

SUMMARY
Purpose – to evaluate the performace of the Oxipit “ChestLink” software in identifying normal chest X-ray studies on “Iconomix” data sample. Oxipit “ChestLink” is an artificial intelligence based software which purpose is to identify and generate reports for a fraction of the normal chest X-ray studies performed in an institution.
Methods – a retrospective sample of anonymized chest X-ray studies from the time period of 2020.01.01-2021.01.01 was selected. The studies were processed by the “ChestLink” solution which identified and generated reports for the studies which the solution classified as normal. The ground truth for the studies was extracted using NLP algorithms on the retrospective final radiologist reports of the studies to extract the structured information about the normal/abnormal radiological findings identified in the study by the radiologist. Using the predictions by the software and the structured ground truth extracted from the final reports we calculated the accuracy metrics. The main metrics used for evaluation were a) sensitivity b) specificity c) normals reporting fraction d) overall reporting fraction.
Results – a total of 4231 from the time period of 2020.01.01-2020.12.31 patients were examined. 3671 (87 %) of these studies were included in the final analysis based on the criteria of 1)final radiologist report is available 2)PA projection 3)Adult patient (age >18). 2743 studies (75 %) of these studies were reported to be normal by radiologists based on the labels extracted from the final reports. The software predicted 906 (33 %) studies to be normal. 903 (99,66 %) of these studies were described by the reporting radiologist as normal, 1 study contained clinically significant radiological findings, 2 studies contained not clinically significant radiological findings. The resulting metrics were:
a) 99,66 % sensitivity
b) 33 % specificity
c) 33 % normals reporting rate
d) 24,67 % overall reporting rate.

Conclusions – the software demonstrated highly sensitive performance in identifying normal chest X-ray studies using a large representative sample from “Iconomix”. The results indicate that AI based software solutions could potentially be used to alleviate the radiologist load by automating the analysis and reporting of a fraction of the normal chest X-ray studies performed in an institution.