A systematic review of artificial intelligence models in ocular tumour diagnosis

Ocular tumours are detrimental to quality of life and survival. The diagnostic capabilities of artificial intelligence (AI) have increased in the past decade. This systematic review aims to evaluate the diagnostic performance of AI models across external, anterior segment, and posterior segment ocular tumours.

A systematic literature search of Ovid Embase, MEDLINE, and the Cochrane Library was performed for studies on AI in ocular tumour diagnosis published from January 2000 to January 2025. Quantitative outcomes were diagnostic accuracy, sensitivity, specificity, and the area under the receiver operating characteristic curve. Findings were synthesized using descriptive statistics. The QUADAS-2 tool assessed the risk of bias and applicability.

Of the 1 947 studies screened, 23 studies were included and categorized as external ocular tumour diagnosis (n = 12), anterior segment tumour diagnosis (n = 2), posterior ocular tumour diagnosis (n = 8), or general ocular tumour diagnosis (n = 1). The main AI models used were DenseNet (n = 4) and ResNet models (n = 3). Weighted mean AI accuracy was 91.4% (81.8% to 98.3%) for external ocular tumour diagnosis, 89.8% (78.1% to 99.0%) for posterior segment ocular tumour diagnosis, and 98.5% for anterior segment tumour diagnosis in the only reporting study. Of the 10 studies comparing AI diagnostic accuracy with physicians, 2 reported significantly higher diagnostic accuracy among ophthalmologists (p < 0.05). Quality assessment demonstrated low or unclear risk of bias and applicability concerns in 69.5% of studies.

AI tools are a potential avenue for efficient and accurate ocular tumour diagnosis. Further studies comparing ophthalmologists' diagnostic performance to AI diagnostic performance are needed.