Can machines see cancer? A systematic review and diagnostic meta-analysis of machine learning in retinoblastoma and leukocoria detection.

Abstract

There is a growing interest in the use of machine learning (ML) for the diagnosis of retinoblastoma and leukocoria, and this study aims to systematically evaluate its performance compared with reference standards. A systematic review and meta-analysis were performed in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) statement. We included studies using ML to diagnose retinoblastoma or leukocoria and providing enough data for analysis of diagnostic accuracy. We calculated sensitivity, specificity, and other measures of diagnostic performance. Twelve studies were included. Pooled sensitivity for retinoblastoma and leukocoria detection was 0.972 with ML models, indicating high potential for screening. However, high heterogeneity in the analyses was observed. The review also noted biases in some studies, along with small sample sizes that would limit generalizability. ML models appear to be promising for retinoblastoma diagnosis; however, limitations in specificity and potential methodological bias need further investigation. Incorporating research that used photographs taken with smartphone cameras indicates that ML-based diagnosis may become even more widely available through the use of such technology. Future studies need to have better specificity of the model, less bias in the methodology, must be conducted on large-scale datasets and they should address the cost-effective analysis compared with traditional methods. The incorporation of ML into the practice of retinoblastoma diagnosis has the capacity to transform the mode of detecting this condition and ultimately enhance patient management.