Automatic detection and measurement system for aortic aneurysms using deep learning-based artificial intelligence.

This study evaluated the performance of a deep learning-based artificial intelligence (DLAI) system for detecting fusiform aortic aneurysms and measuring the diameters of the entire aorta on non-contrast CT images.We retrospectively collected 160 non-contrast CT images (130 males, 71.8 ± 9.2 years) to evaluate aortic aneurysms and used them as training data. To evaluate the accuracy of the DLAI system for the automatic detection and measurement of aortic aneurysms, another 190 non-contrast CT images (153 males, 70.7 ± 10.3 years) were collected and used as validation data. First, the accuracy of DLAI results was calculated by referring to the radiology reports. An expert radiologist read the test data again and calculated the final accuracy by referring to the DLAI predictions.Dice scores for aortic segmentation were 0.90, 0.94, 0.93, and 0.84 for the entire aorta, thoracic aorta, abdominal aorta, and iliac artery, respectively. The sensitivity, positive predictive value, and F-measure of the DLAI system for detecting aortic aneurysms were 0.81, 0.83, and 0.82, respectively. Eleven (4.4%) aneurysms were newly identified after the radiologist's second reading based on DLAI predictions. The final accuracy of the DLAI detection increased to 0.83, 0.87, and 0.85 for sensitivity, positive predictive value, and F-measure, respectively. The size error in the measurement of aortic aneurysms between the DLAI system and radiology reports was 0.86 ± 2.72 mm, with a strong correlation (ICCs = 0.97).Our DLAI system was highly accurate in detecting aneurysms and measuring the diameters of the entire aorta on non-contrast CT images.