Glaucoma is the mainly cause of irreversible blindness characterized with structural changes of optic nerve head, retinal ganglion cell death and loss of visual acuity, with estimated 64.3 million patients aged from 40 to 80 worldwide. When patients seek medical advice because of visual acuity, glaucoma often has advanced to late stage. One major challenge is to identify large number of undiagnosed patients. The reason why screening programs are not employed regularly is that misdiagnoses can be made due to large amounts of false positives in the procedure.Artificial intelligence (AI) has brought new breakthroughs for automated screening for glaucoma，which can perform a task by classifying and identifying data without human intervention, including supervised and unsupervised machine learning two categories. Due to the limitation of manually designed features and disability in detection for high dimensionality of optic disc images, advances in techniques for evaluation of glaucoma came with the use of segmentation‐free approaches. We reported a deep learning system and assessed its generalizability in various data sets, with similar patterns in sensitivity (82.2%–96.1%) and specificity (70.4%–97.1%).Visual field data presents low dimensionality and high noise compared to optic disc images, whose hidden patterns can be detected by unsupervised learning better. Two most reported unsupervised algorithms were clustering and component analysis. Such technologies adopted in glaucoma screening have significant potential to increase accessibility in remote areas and decrease the cost for widely screening. Further studies similarly demonstrated that machine learning can perform comparable to or over human experts in mean deviation, pattern standard deviation and glaucoma hemifield test. A back‐propagation neural network was even reported to successfully detect visual field progression with Area Under Curve (AUC) of 0.92. Further advances were also made through algorithms based on CNN showing both higher sensitivity and specificity than traditional machine learning method.In terms of optical coherence tomography (OCT), early studies based on MLC using TD‐OCT showed an accuracy not worse than non‐artificial intelligence analysis. The latest OCT technologies, including SD‐OCT and SS‐OCT, combined with deep learning are reported more sensitive for early glaucoma detection, with a high AUC up to 0.937. Recently a study reported the ability of their model to detect RNFL thickness from OCT with the AUC of 0.944. Recent years, more parameters have been put in including OCT RNFL thickness, standard automated perimetry and confocal scanning laser ophthalmoscopy imaging, making an improvement in glaucoma detection.However, there remains several limitations and further studies is required. First, the exact mechanism how DL algorithms evaluate features and make predictions remains unknown, which is called ‘black boxes’. Another limitation is that it can be difficult for DL to classify glaucoma in eyes with less severe disease manifestations or multiple comorbid eye conditions, especially high myopia, which asks for a large images database. Furthermore, in order to develop a more dependent screening method, other clinical parameters including intraocular pressure and central corneal thickness should be integrated, and application and validation of such in a real‐world screening need additional research to support.
Read full abstract
Oct 23, 2024
Selma Meşen + 2
Open Access