Evaluation des performances diagnostiques d'un logiciel d'intelligence artificielle sur une cohorte de nodules pulmonaires malins de tailles différentes
Langue Français
Langue Français
Auteur(s) : Roth Victor
Directeur : Mickaël Ohana
Composante : MEDECINE
Date de création : 30-06-2022
Description : Médecine (radiologie et imagerie médicale), Objective: To evaluate the diagnostic performance of a Deep Learning software (InferRead CT Lung, Infervision, China) predictive of malignancy in a cohort of "extreme" pulmonary nodules of different sizes, all malignant. The secondary objective was to assess the possible impact of nodule characteristics on the malignancy score.Material and method: retrospective, single-center study including malignant pulmonary nodules of size ≤ 30 mm biopsied between December 1, 2019, and April 30, 2021, with CT imaging less than 3 months old. Anonymized images were sent to a local server for analysis by the artificial intelligence software. According to the probability of malignancy given by the software, nodules were classified into 2 groups: - 50% = "safe" and ≥ 50% = "suspicious." To analyze the performance of the software, nodules were separated into several subgroups depending on their size: -6 mm, 6 to 8 mm, 9 to 10 mm, 11 to 15 mm, and -15 mm. For each of these size categories, the "suspect" subgroup was considered a correct software result and the other subgroups were considered failures to analyze. Results: 137 pulmonary nodules were included in 97 patients. Of all the nodules included, 108 were correctly detected and classified as "suspicious". The correct response rate of the AI software was 0.79. According to the different size categories, the rate is: 0.07 for nodules - 6 mm, 0.44 for nodules from 6 to 8 mm, 0.77 for nodules from 9 to 10 mm, 0.94 for nodules from 11 to 15 mm and 0.97 for nodules - 15 mm. Among the characteristics of the nodules, only their size had a significant impact on the malignancy score (p - 0.00001). Conclusion: This retrospective monocentric study, on a cohort of 137 "extreme" nodules, all malignant and of variable size, shows that the performance of an AI characterization software varies greatly according to nodule size. If a high specificity with a high NPV remains likely for nodules larger than 10mm, the number of false negatives for nodules ≤10mm remains a concern and probably does not allow to consider the use of the software as a "gatekeeper" to surpass the follow-up of 6-10mm nodules., Thèses et écrits académiques
Mots-clés libres : Intelligence artificielle, Deep-learning, 616.075, Nodules pulmonaires malins, Cancer du poumon, Machine lear, Apprentissage profond, Performances diagnostiques
Couverture : FR
Directeur : Mickaël Ohana
Composante : MEDECINE
Date de création : 30-06-2022
Description : Médecine (radiologie et imagerie médicale), Objective: To evaluate the diagnostic performance of a Deep Learning software (InferRead CT Lung, Infervision, China) predictive of malignancy in a cohort of "extreme" pulmonary nodules of different sizes, all malignant. The secondary objective was to assess the possible impact of nodule characteristics on the malignancy score.Material and method: retrospective, single-center study including malignant pulmonary nodules of size ≤ 30 mm biopsied between December 1, 2019, and April 30, 2021, with CT imaging less than 3 months old. Anonymized images were sent to a local server for analysis by the artificial intelligence software. According to the probability of malignancy given by the software, nodules were classified into 2 groups: - 50% = "safe" and ≥ 50% = "suspicious." To analyze the performance of the software, nodules were separated into several subgroups depending on their size: -6 mm, 6 to 8 mm, 9 to 10 mm, 11 to 15 mm, and -15 mm. For each of these size categories, the "suspect" subgroup was considered a correct software result and the other subgroups were considered failures to analyze. Results: 137 pulmonary nodules were included in 97 patients. Of all the nodules included, 108 were correctly detected and classified as "suspicious". The correct response rate of the AI software was 0.79. According to the different size categories, the rate is: 0.07 for nodules - 6 mm, 0.44 for nodules from 6 to 8 mm, 0.77 for nodules from 9 to 10 mm, 0.94 for nodules from 11 to 15 mm and 0.97 for nodules - 15 mm. Among the characteristics of the nodules, only their size had a significant impact on the malignancy score (p - 0.00001). Conclusion: This retrospective monocentric study, on a cohort of 137 "extreme" nodules, all malignant and of variable size, shows that the performance of an AI characterization software varies greatly according to nodule size. If a high specificity with a high NPV remains likely for nodules larger than 10mm, the number of false negatives for nodules ≤10mm remains a concern and probably does not allow to consider the use of the software as a "gatekeeper" to surpass the follow-up of 6-10mm nodules., Thèses et écrits académiques
Mots-clés libres : Intelligence artificielle, Deep-learning, 616.075, Nodules pulmonaires malins, Cancer du poumon, Machine lear, Apprentissage profond, Performances diagnostiques
Couverture : FR
Type : Thèse d'exercice, ressource électronique
Format : Document PDF, Document PDF
Source(s) :
Format : Document PDF, Document PDF
Source(s) :
- http://www.sudoc.fr/266748937
Entrepôt d'origine :
Identifiant : ecrin-ori-337281
Type de ressource : Ressource documentaire
Identifiant : ecrin-ori-337281
Type de ressource : Ressource documentaire