Artificial intelligence and machine learning in ocular oncology, retinoblastoma (ArMOR)

Le résumé fourni par la source

PURPOSE: To test the accuracy of a trained artificial intelligence and machine learning (AI/ML) model in the diagnosis and grouping of intraocular retinoblastoma (iRB) based on the International Classification of Retinoblastoma (ICRB) in a larger cohort. METHODS: Retrospective observational study that employed AI, ML, and open computer vision techniques. RESULTS: For 1266 images, the AI/ML model displayed accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) of 95%, 94%, 98%, 99%, and 80%, respectively, for the detection of RB. For 173 eyes, the accuracy, sensitivity, specificity, PPV, and NPV of the AI/ML model were 85%, 98%, 94%, 98%, and 94% for detecting RB. Of 173 eyes classified based on the ICRB by two independent ocular oncologists, 9 (5%) were Group A, 32 (19%) were Group B, 21 (12%) were Group C, 37 (21%) were Group D, 38 (22%) were Group E, and 36 (21%) were classified as normal. Based on the ICRB classification of 173 eyes, the AI/ML model displayed accuracy, sensitivity, specificity, PPV, and NPV of 98%, 94%, 99%, 94%, and 99% for normal; 97%, 56%, 99%, 71% and 98% for Group A; 95%, 75%, 99%, 96%, and 95% for Group B; 95%, 86%, 96%, 75%, and 98% for Group C; 92%, 76%, 96%, 85%, and 94% for Group D; and 94%, 100%, 93%, 79%, 100% for Group E, respectively. CONCLUSION: These observations show that expanding the image datasets, as well as testing and retesting AI models, helps identify deficiencies in the AI/ML model and improves its accuracy.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.