Szczegóły publikacji
Opis bibliograficzny
Artificial intelligence-based algorithm for stent coverage assessments / Joanna Fluder-Włodarczyk, Mikhail DARAKHOVICH, Zofia SCHNEIDER, Magda Roleder-Dylewska, Magdalena Dobrolińska, Tomasz Pawłowski, Wojciech Wojakowski, Paweł Gasior, Elżbieta POCIASK // Journal of Personalized Medicine [Dokument elektroniczny]. — Czasopismo elektroniczne ; ISSN 2075-4426 . — 2025 — vol. 15 iss. 4 art. no. 151, s. 1–12. — Wymagania systemowe: Adobe Reader. — Bibliogr. s. 10–12, Abstr. — Publikacja dostępna online od: 2025-04-11
Autorzy (9)
- Fluder-Włodarczyk Joanna
- Darakhovich Mikhail
- AGHSchneider Zofia
- Roleder-Dylewska Magda
- Dobrolińska Magdalena
- Pawłowski Tomasz
- Wojakowski Wojciech
- Gasior Paweł
- AGHPociask Elżbieta
Słowa kluczowe
Dane bibliometryczne
| ID BaDAP | 160051 |
|---|---|
| Data dodania do BaDAP | 2025-06-25 |
| Tekst źródłowy | URL |
| DOI | 10.3390/jpm15040151 |
| Rok publikacji | 2025 |
| Typ publikacji | artykuł w czasopiśmie |
| Otwarty dostęp |  |
| Creative Commons | |
| Czasopismo/seria | Journal of Personalized Medicine |
Abstract
Background: Neointimal formation after stent implantation is an important prognostic factor since delayed healing may lead to stent thrombosis. In vivo, optical coherence tomography (OCT) can most precisely assess stent strut coverage. Since analyzing neointimal coverage is time-consuming, artificial intelligence (AI) may offer valuable assistance. This study presents the preliminary results of the AI-based tool’s performance in detecting and categorizing struts as covered and uncovered. Methods: The algorithm was developed using the YOLO11 (You Only Look Once) neural networks. The first step was preprocessing, then data augmentation techniques were implemented, and the model was trained. Twenty OCT pullbacks were used during model training, and two OCT pullbacks were used in the final validation. Results: The presented tool’s performance was validated against two analysts’ consensus. Both analysts showed moderate intraobserver agreement (κ = 0.57 for analyst 1 and κ = 0.533 for analyst 2) and fair agreement with each other (κ = 0.389). The algorithm’s detection of struts was satisfactory (a 92% positive predictive value (PPV) and a 90% true positive rate (TPR)) and was more accurate in recognizing covered struts (an 81% PPV and an 85% TPR) than uncovered struts (a 73% PPV and a 60% TPR). The agreement was κ = 0.444. Conclusions: The initial results demonstrated a good detection of struts with a more challenging uncovered strut classification. Further clinical studies with a larger sample size are needed to improve the proposed tool.
