Introduction: Geographic atrophy (GA), an advanced form of age-related macular degeneration, is a significant cause of blindness and represents one of the major therapeutic challenges in ophthalmology. Optical coherence tomography (OCT) is an advanced retinal imaging technology capable of identifying specific structural changes, which may serve as potential biomarkers in various retinal diseases. Recent studies in the literature have shown that small intraretinal hyperreflective foci (HRF), less than 30 μm in diameter with intermediate reflectivity comparable to that of the retinal nerve fiber layer, are aggregates of activated microglial cells and thus potential biomarkers of pathology. HRF have recently been identified and quantified in eyes affected by GA. Their quantity varies depending on the different phenotypes of GA, suggesting a possible role for retinal neuroinflammation in the development and progression of GA. Manual counting of HRF is time-consuming and requires experienced examiners, highlighting the need for alternative methods to quantify HRF for potential clinical use. Study Aim:The purpose of this study is to evaluate and compare the accuracy of two methods for quantifying HRF in the context of GA, by comparing an artificial intelligence (AI) algorithm with manual counting performed by an experienced examiner. Materials and Methods: High-resolution OCT scans of patients with different forms of GA (bilateral GA or unilateral GA with choroidal neovascularization in the contralateral eye) and age-matched healthy controls were analyzed. The images were processed using an artificial intelligence algorithm (Ophthal) for the automatic quantification of HRF, which has been validated in patients with diabetic retinopathy. The same OCT images were also analyzed by an experienced examiner who manually identified and quantified the HRF. Results and Discussion: A total of 38 eyes from 29 patients with bilateral GA (B-GA Group), 19 eyes from 19 patients with unilateral GA (U-GA Group), and 19 eyes from 18 healthy controls were analyzed. The number of HRF manually measured was found to be significantly lower compared to the AI measurements in all study groups (p<0.0001 for B-GA, p=0.0429 for U-GA, and p=0.0288 for controls), with a mean difference of 23.7±11.8 foci in the B-GA group, 10.5±18.2 in the U-GA group, and 6.2±2.3 in the control group. After applying a statistical "adjustment" model accounting for central retinal thickness and foveal sparing, the mean difference in HRF counts between the two methods was 3.1±11.1 in the B-GA group (p=0.1346) and 14.5±18.1 in the U-GA group (p=0.0026). The entire group of GA patients was also analyzed, showing a mean difference of 19.4±31.2 foci, which was reduced to 5.6±22.3 foci (p=0.328) after applying the statistical "adjustment" model. Conclusion: The AI software showed significant discrepancies in detecting and counting HRF in GA compared to manual counting, suggesting that AI systems are not easily or automatically applicable to retinal diseases other than those for which they were specifically designed and trained. In geographic atrophy, the retinal structural changes characteristic of the disease complicate the application of this AI system, especially in certain specific GA phenotypes.
Introduzione: L’atrofia geografica (GA), forma evoluta della maculopatia legata all’età, è causa importante di cecità e costituisce una delle maggiori sfide in ambito terapeutico nel campo dell’oftalmologia. La tomografia a coerenza ottica (OCT) rappresenta una tecnologia per immagini retinica avanzata in grado di indentificare alcune specifiche alterazioni strutturali, potenziali biomarcatori in diverse patologie retiniche. Studi recenti in letteratura hanno evidenziato come piccoli foci iperriflettenti intraretinici (HRF) all’OCT, con un diametro inferiore a 30 μm e una riflettività intermedia, comparabile a quella dello strato delle fibre nervose retiniche costituiscono aggregati di cellule microgliali attivate, e quindi possibili biomarcatori di patologia. HRF sono stati recentemente identificati e quantificati in occhi affetti da GA. La loro numerosità cambia in relazione al diverso fenotipo di GA, indicando in questi un possibile diverso ruolo della neuroinfiammazione retinica nello sviluppo e progressione della GA. Il processo di conteggio manuale richiede tempo ed esaminatori esperti, rendendo necessaria la ricerca di metodi alternativi per la quantificazione dei HRF e quindi un possibile loro utilizzo nella pratica clinica. Scopo dello studio: Il presente studio si propone di valutare e confrontare l’accuratezza di due metodologie di quantificazione degli HRF nel contesto della GA, confrontando un algoritmo di intelligenza artificiale (IA) con il conteggio manuale eseguito da un esaminatore esperto. Materiali e metodi: Sono state analizzate scansioni OCT ad alta risoluzione di pazienti affetti da differenti forme di GA (GA bilaterale o GA monolaterale con neovascolarizzazione coroideale nell’occhio adelfo) e di soggetti sani di età confrontabile con la popolazione affetta. Le immagini sono state elaborate utilizzando un algoritmo di intelligenza artificiale (Ophthal) per la quantificazione automatica degli HRF validato in pazienti affetti da retinopatia diabetica. Le stesse immagini OCT sono state analizzate da un esaminatore esperto che ha manualmente identificato e quantificato gli HRF. Risultati e discussione: Sono stati quindi analizzati 38 occhi di 29 pazienti con GA bilaterale (Gruppo B-GA), 19 occhi di 19 pazienti con GA unilaterale (Gruppo U-GA) e 19 occhi di 18 soggetti sani (Controlli).Il numero di HRF misurati manualmente si è rivelato significativamente inferiore rispetto a quanto misurato dall’ IA in tutti i gruppi di studio (p<0.0001 per B-GA, p=0.0429 per U-GA e p=0.0288 per controlli), con una differenza media di 23.7±11.8 foci nel gruppo B-GA, 10.5 ± 18.2 nel gruppo U-GA, e di 6.2±2.3 nei controlli. Dopo applicazione di un modello statistico di “aggiustamento” che ha tenuto in considerazione spessore retinico centrale e risparmio foveale, la differenza media tra le due metodiche nel numero di HRF è stata di 3.1 ± 11.1 nel gruppo B-GA (p=0.1346), e di 14.5±18.1 nel gruppo U-GA (p=0.0026). Si è inoltre considerato il gruppo totale dei soggetti con GA, i quali presentavano una differenza media di 19.4±31.2 foci, ridotta a 5.6 ± 22.3 foci (p=0.328) in seguito ad applicazione del modello statistico di “aggiustamento”. Conclusione: Il software di IA ha dimostrato significative discrepanze nella rilevazione nel conteggio degli HRF nella GA rispetto al loro conteggio manuale, indicando come software di IA non siano di facile ed automatica applicazione in patologie retiniche diverse da quelle per i quali sono stati ideati e allenati. Nell’atrofia geografica, alterazioni strutturali retiniche che caratterizzano tale patologia rendono l’applicazione di tale sistema di IA complessa, soprattutto in alcuni specifici fenotipi di GA.
Analisi di foci iperriflettenti retinici nell'atrofia geografica
MAGGIOLO, ALBERTO
2023/2024
Abstract
Introduction: Geographic atrophy (GA), an advanced form of age-related macular degeneration, is a significant cause of blindness and represents one of the major therapeutic challenges in ophthalmology. Optical coherence tomography (OCT) is an advanced retinal imaging technology capable of identifying specific structural changes, which may serve as potential biomarkers in various retinal diseases. Recent studies in the literature have shown that small intraretinal hyperreflective foci (HRF), less than 30 μm in diameter with intermediate reflectivity comparable to that of the retinal nerve fiber layer, are aggregates of activated microglial cells and thus potential biomarkers of pathology. HRF have recently been identified and quantified in eyes affected by GA. Their quantity varies depending on the different phenotypes of GA, suggesting a possible role for retinal neuroinflammation in the development and progression of GA. Manual counting of HRF is time-consuming and requires experienced examiners, highlighting the need for alternative methods to quantify HRF for potential clinical use. Study Aim:The purpose of this study is to evaluate and compare the accuracy of two methods for quantifying HRF in the context of GA, by comparing an artificial intelligence (AI) algorithm with manual counting performed by an experienced examiner. Materials and Methods: High-resolution OCT scans of patients with different forms of GA (bilateral GA or unilateral GA with choroidal neovascularization in the contralateral eye) and age-matched healthy controls were analyzed. The images were processed using an artificial intelligence algorithm (Ophthal) for the automatic quantification of HRF, which has been validated in patients with diabetic retinopathy. The same OCT images were also analyzed by an experienced examiner who manually identified and quantified the HRF. Results and Discussion: A total of 38 eyes from 29 patients with bilateral GA (B-GA Group), 19 eyes from 19 patients with unilateral GA (U-GA Group), and 19 eyes from 18 healthy controls were analyzed. The number of HRF manually measured was found to be significantly lower compared to the AI measurements in all study groups (p<0.0001 for B-GA, p=0.0429 for U-GA, and p=0.0288 for controls), with a mean difference of 23.7±11.8 foci in the B-GA group, 10.5±18.2 in the U-GA group, and 6.2±2.3 in the control group. After applying a statistical "adjustment" model accounting for central retinal thickness and foveal sparing, the mean difference in HRF counts between the two methods was 3.1±11.1 in the B-GA group (p=0.1346) and 14.5±18.1 in the U-GA group (p=0.0026). The entire group of GA patients was also analyzed, showing a mean difference of 19.4±31.2 foci, which was reduced to 5.6±22.3 foci (p=0.328) after applying the statistical "adjustment" model. Conclusion: The AI software showed significant discrepancies in detecting and counting HRF in GA compared to manual counting, suggesting that AI systems are not easily or automatically applicable to retinal diseases other than those for which they were specifically designed and trained. In geographic atrophy, the retinal structural changes characteristic of the disease complicate the application of this AI system, especially in certain specific GA phenotypes.File | Dimensione | Formato | |
---|---|---|---|
Maggiolo_Alberto (PDFA).pdf
accesso riservato
Dimensione
1.28 MB
Formato
Adobe PDF
|
1.28 MB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/78415