Análise da tomografia e biomecânica da córnea com o uso de inteligência artificial para o diagnóstico e prevenção da ectasia iatrogênica após cirurgia de correção visual a laser

• Autor(a) principal: Lopes, Bernardo Teixeira [UNIFESP]
• Data de Publicação: 2018
• Tipo de documento: Tese
• Idioma: por
• Título da fonte: Repositório Institucional da UNIFESP
• Texto Completo: https://sucupira.capes.gov.br/sucupira/public/consultas/coleta/trabalhoConclusao/viewTrabalhoConclusao.jsf?popup=true&id_trabalho=6773000; https://repositorio.unifesp.br/handle/11600/52901
• topic: Refractive surgery; Iatrogenic ectasia; Keratoconus; Artificial intelligence; Corneal tomography and biomechanics; Cirurgia refrativa; Ectasia Iatrogênica; Ceratocone; Inteligência artificial; Tomografia e biomecânica da córnea

Purposes: To test corneal geometric (tomographic) and biomechanical parameters and to optimize their combination in the diagnosis of keratoconus and to quantify the risk of progressive ectasia after laser visual correction surgery. Methods: In the first study, a prospective review of the accuracy of tomographic indices obtained with rotational Scheimpflug system (Pentacam; Oculus Optikgeräte, Inc., Wetzlar, Germany) in the diagnosis of keratoconus, so as the identification of subclinical ectasia cases, defined as the contralateral eyes with normal topography from patients with asymmetric ectasia clinically detectable in the other eye. In the second study, we have developed models to retrospectively compare the tomographic data of the preoperative status of patients undergoing LASIK, with documented stability of at least two years and of the preoperative status of patients who developed progressive ectasia after LASIK. These models were tested for external validation including the preoperative status of two stable populations and in cases of keratoconus and cases with subclinical ectasia (normal topography eye of patients with asymmetric ectasia). In this study, the PRFI(Pentacam Random Forest Index) was developed from the artificial intelligence (AI) models to optimize the separation between the groups. In the third study, the precision (repeatability) and reproducibility study of Corvis ST measurements (Oculus Optikgeräte, Inc., Wetzlar, Germany) were tested in normal patients in three different devices, with three consecutive measures in each device. In the fourth study, we have evaluated the special distribution of corneal thickness in its horizontal section, in order to establish a pachymetric model with AI methods to increase the accuracy of the diagnosis of keratoconus. In the fifth study, we have evaluated and combined deformation indices and the horizontal pachymetry to detect keratoconus, resulting in the development of a logistic regression model (CBI – Corvis Biomechanical Index). In the sixth study, the biomechanical parameters were combined with the tomographic data with AI models, with the aim of increasing the accuracy in the diagnosis of ectasia, including subclinical cases (eye with normal topography in very asymmetric cases of ectasia). In this study has been developed the TBI (Tomography Biomechanical Index), which was tested in the seventh study to external validation. Results: The tomographic indices involving both surfaces of the cornea and the pachymetric map have presented superior capacity when compared with indices based exclusively on the anterior surface. In the second study, we have observed a higher accuracy of the PRFI in susceptible cases (preoperative of patients who developed progressive postoperative ectasia) and cases of asymmetric ectasia (AUC: 0.966 and 0.968, respectively) when compared to the best existing tomographic index (BADD [BelinAmbrósio Deviation Index] AUC: 0.845 and 0.893, respectively). In the third study, the indices of corneal deformation and pressure measurements (IOP) have presented acceptable repeatability and reproducibility in normal patients. In the fourth study, we have observed that the AI analysis of the horizontal pachymetric profile can increase the accuracy in the detection of keratoconus when comparing with the central thickness. In the fifth study, the CBI, with the combination of deformation parameters and horizontal pachymetric profile data has presented acceptable accuracy to the diagnosis of keratoconus (AUC = 0.983). The sixth study has shown that the integration of tomography data with corneal deformation data increased the accuracy in identifying the clinical ectasia (AUC = 1) and the subclinical cases (asymmetric ectasia; AUC = 0.985). The seventh study has demonstrated the robustness of the TBI with external validation. Conclusions: The diagnosis of corneal ectasia has evolved with the characterization of tomography (threedimensional geometry study) and biomechanics (the study of the deformation) with Scheimpflug images. The analysis with AI models can considerably increase accuracy in subclinical cases, said to be susceptible to develop ectasia. The parameters that characterize the biomechanical properties in vivo present acceptable repeatability and reproducibility. Statistical and AI methods are important to the applicability and clinical decision of diagnostic methods.