Purchase this article with an account.
Kaveri Thakoor, Ari Leshno, Sol La Bruna, Emmanouil Tsamis, Gustavo De Moraes, Paul Sajda, Noga Harizman, Jeffrey M Liebmann, George A Cioffi, Donald C Hood; Evaluation of a Deep Learning Model on a Real-World Clinical Glaucoma Dataset. Invest. Ophthalmol. Vis. Sci. 2022;63(7):2026 – A0467.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
To evaluate a deep learning model (DLM) for detecting glaucoma based upon OCT probability maps by applying it to eyes from a university-based clinical practice. We asked if the DLM diagnosis accurately reflects the diagnosis made by OCT experts using a full OCT report, and if it had the potential to aid in the clinical diagnosis.
The model: A DLM  was previously trained and validated on OCT retinal nerve fiber layer probability maps (RNFL p-maps) (Fig. 1A) from good quality scans of patients without obvious co-morbidities or extreme refractive errors. The real-world data: OCT scans were obtained from 99 eyes from 59 patients visiting a university-based, glaucoma practice. 4 eyes with unreadable scans were excluded. Unlike the eyes used to train the DLM, the 95 eyes were older (55 eyes, aged 70 to 98 years), with a range of disease severity, and various co-morbidities (e.g., ERM, AMD, high myopia). The analysis: The DLM output ranges from 0% (H: healthy) to 100% (G: optic neuropathy consistent with glaucoma). Eyes were categorized as H (≤5%), G (≥95%), or UNC (uncertain), and compared to the grading of a Hood report (Fig. 1B) by 4 OCT experts (OCT-E), who rated each eye H, G, or UNC. G and UNC judgments were combined, as both require follow-up. To assess the potential value of the DLM, a glaucoma expert made a clinical decision (CD) of H, G, or UNC (suspect) twice for each eye, first (CD1) based upon past chart notes, visual fields, and OCT scans, except for the recent Hood report, and again (CD2) to see if the OCT Hood report with the DLM output and heatmap (Fig. 1B, 1C) would alter CD1.
The DLM decision agreed with the OCT-E decision in 95% of the eyes; all 5 disagreements involved an UNC judgment. Of the 9 eyes with a CD1 of UNC (suspect), CD2 changed to G in 3 and H in 3, in agreement with the DLM. Of the 7 eyes with a CD1 of H, but a DLM grade of G, CD2 changed to “probably G” in 4 (e.g., Fig. 1B) and UNC in 2. Of the 6 eyes with a CD1 of G, but a DLM grade of H, CD2 changed to H in 1, and in 4 eyes, the CD2 noted that other factors such as ERMs or retinal disease probably contributed.
The DLM output based upon the RNFL p-map showed excellent agreement with the OCT-E decisions based upon the complete OCT report (Fig. 1B). A post-hoc analysis suggested that the DLM has the potential to aid in clinical diagnosis.  Thakoor et al., IEEE TBME, 2021.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.
This PDF is available to Subscribers Only