Computer vision programs in medicine: dataset development, classification, and clinical use

Abstract

Objective: To develop and systematize a methodology for constructing datasets for computer vision programs (CVPs) in clinical medicine; to propose and substantiate a multicriteria classification of CVPs and to characterize their clinical use at three organizational levels: the patient, the clinician and the medical organization.

Material and methods. The study systematizes the concepts of "dataset" and "database" in medical artificial intelligence, develops a multicriteria classification of CVPs, and , and characterizes their clinical use. The study employed terminological and comparative analyses, as well as a and classification approach.

Results. The concepts of "dataset" and "database" were clarified, and their key characteristics were formalized. A six-stage methodology for constructing a dataset from a clinical database was described, comprising extraction, de-identification, verification, structuring, balancing, and documentation. A four-criteria classification of CVPs was proposed and substantiated, based on input-data modality, number of models used, deployment setting, and integration with clinical equipment. The clinical application of CVPs was characterized at three organizational levels: the patient, the clinician, and the medical organization.

Conclusion. Proper dataset construction is a key determinant of the clinical validity of CVPs. The proposed classification provides a standardized description of these systems, enabling their valid comparison and regulatory evaluation. The identification of the levels at which CVP is used in clinical practice helps optimize its integration into real-world healthcare settings.