The Promise and Pitfalls: A Systematic Review of AI-Powered Diagnostics in Modern Oncology

Abstract

Purpose: Despite the rapid expansion of AI-powered diagnostics in oncology, most existing literature reports algorithm performance in isolated, retrospective settings, and no recent systematic review has jointly synthesised algorithm types, validation rigor, implementation barriers, and patient-outcome evidence. This leaves a gap in understanding how far AI diagnostics have actually progressed from development toward clinical use. This review addresses that gap by examining AI-powered oncology diagnostics across four areas: the types of AI algorithms developed, the validation techniques employed, the challenges and enablers of clinical implementation, and the evidence for improved patient outcomes.
Methodology: Following PRISMA guidelines, a systematic search of Scopus, BioMed Central, and Google Scholar identified peer-reviewed empirical studies published between 2020 and 2025. Of 1,265 records screened, 27 studies met inclusion criteria. Data extraction covered algorithm type, cancer type, dataset, validation technique, performance metrics, and implementation status; QUADAS-2 was used to assess risk of bias.
Results: Most studies used deep learning models, particularly CNNs, for imaging and histopathology diagnosis, with accuracies frequently exceeding 90% and AUCs up to 0.987; breast and lung cancers were most studied. However, validation was predominantly internal and retrospective, with only ~15% of studies reaching clinical trials or real-world implementation. Poor interpretability, limited external validation, and integration barriers constrain clinical adoption, alongside moderate-to-high risk of bias in patient selection and reference standards.
Novelty and Contribution: This review uniquely maps the gap between high-performing AI models and limited clinical adoption, offering a roadmap bridging development and implementation.
Implications: Interdisciplinary collaboration, prospective multicenter trials, and equitable, ethical integration frameworks are needed to advance clinical AI adoption in oncology.