Auteur(s) : Hurstel Florie
Composante : MEDECINE
Date de création : 30-06-2021

Description : Médecine (radiodiagnostic et imagerie médicale), Purpose : To evaluate the diagnostic performances of a commercially available AI system for breast cancer detection in different configurations (stand-alone AI, unaided radiologist, AI-aided radiologist). Materials et Methods: A retrospective, monocentric, multireader, comparative study, with pathological gold standard, was performed. 498 digital mammogramms from 249 women aged 40-80 years were included: 150 patients with normal mammograms and 99 patients with malignant lesions. Three reading groups were compared: stand-alone AI, unaided radiologist, and AI-aided radiologist. Results: The AUC of radiologists (0.89, CI95=0.86 - 0.93) was higher than the AUC of stand-alone AI (0.80, CI95=0.7-0.9). The AUC of AI-aided radiologists (0.91, CI95=0.87 - 0.94) was higher than the AUC of unaided radiologists, though the difference was not statistically significant. Despite stand-alone AI producing a significant number of false positives, they were largely negated by human readers, leading to only a small decrease in specificity for the AI+radiologist reading (0.95 [0.92 - 0.97] vs 0.96 [0.94 - 0.98]). Inter-rater reliability tended to improve with the use of AI (kappa 0.77 [0.70, 0.83] vs 0.74 [0.66, 0.81]). Conclusion: In this independent cohort, the use of AI allowed a moderate gain in sensitivity. Specificity was only slightly decreased with the use of AI. AI-aided reading has the potential to reduce inter-rater reliability, although larger clinical studies need to be conducted to address this question
Mots-clés libres : Sein -- Cancer, Intelligence artificielle en médecine, Mammographie, 616.075
Couverture : FR