Uveal Melanoma Ground Truth Labeling in Machine Learning

Uveal melanoma (UM) is the most common primary intraocular malignant tumor among adults and has a high risk of metastasis. Recently, artificial intelligence (AI) tools have been developed to support the management of UM across different clinical tasks. The definition of ground truth, the reference standard that models use in training and development, greatly influences the performance and clinical relevance of the models. Currently, there is limited consensus regarding which ground truth methods are most appropriate for each clinical application. This review aims to evaluate the advantages and limitations of available ground truth options in UM and proposes task-specific recommendations based on clinical utility, feasibility, and cost.

A narrative review of the existing literature was conducted to identify and evaluate commonly used ground truth methods for UM AI applications based on factors such as time, cost, invasiveness, and required level of expertise.

Each ground truth method offers distinct benefits and drawbacks in relation to biological precision, invasiveness, availability, cost, and turnaround time. No single ground truth is universally optimal across all applications. Instead, the ideal choice depends on the intended clinical task, and practical alternatives exist to mitigate the constraints that result from limited time and institutional resources.

The selection of ground truth for AI models in UM should be chosen based on the specific clinical task to balance predictive relevance with feasibility of implementation. The adoption of task-specific ground truth standards may improve the development of clinically meaningful AI tools and facilitate their integration into real-world practice.