Role of artificial intelligence in ocular tumors: A systematic review.
journal contribution
posted on 2024-07-31, 01:45authored byShadi Farabi Maleki, Milad Yousefi, Zanyar Hajiesmailpoor, Ali Jafarizadeh, Siamak Pedrammehr, Roohallah AlizadehsaniRoohallah Alizadehsani, Juan Manuel Gorriz Saez
e15070 Background: Integrating oncology and artificial intelligence (AI) shows significant potential in improving treatments and early detection of ocular tumors. This systematic review aims to strengthen existing evidence regarding the application of AI in diagnosing and treating ocular tumors. Methods: We systematically searched PubMed, Embase, Web of Science, Scopus, and Google Scholar until January 15, 2024. Inclusion criteria comprised original peer-reviewed human studies (retrospective or prospective) investigating AI approaches in patients with ocular tumors such as retinoblastoma, uveal melanoma, conjunctival melanoma, and choroidal melanoma. The quality of the literature was assessed using JBI checklists. The eligibility criteria involved assessing studies that explored the application of AI in ocular tumor detection, classification, and treatment planning. We employed a rigorous screening process, and subsequent data extraction focused on key characteristics, including tumor types, sample sizes, AI input data specifications, and study aims. Exclusion criteria encompassed studies not meeting the inclusion criteria, lacking relevant outcome measures, exhibiting insufficient methodological quality, or inadequate reporting. Results: The initial database search yielded 903 records, of which 37 studies (retrospective = 24 and prospective = 13) were included. Most studies selected the retinoblastoma (n = 12) and uveal melanoma (n = 14). Other studies utilized other types of ocular tumors, including ocular metastasis (n = 2), ocular surface squamous neoplasia (n = 2), and ocular lymphoma (n = 2). The conjunctival melanoma, choroidal melanoma, intraocular melanoma, cavernous hemangioma, and schwannoma patients were utilized once. The AI-based models were employed for the classification (n = 22), prognosis prediction (n = 6), lesion segmentation (n = 5), and predicting the gene expression (n = 4) tasks. The data types used Included magnetic resonance imaging (MRI), computed tomography (CT), whole-slide imaging (WSI), clinical data, and fundus images. A meta-analysis was precluded because of the significant heterogeneity observed across the included studies in terms of demographics, study design, interventions, and outcome measures. Conclusions: Although imaging and patient presentation have played a crucial role in ocular tumor diagnostics, AI presents promising results in prediction, prognosis, and disease management. Meanwhile, AI's primary role has been in classification, demonstrating significant success. Additionally, continued exploration of less-discussed eye tumors is essential for enhancing AI's predictive and therapeutic capabilities. The absence of reported evaluation criteria and low evaluation parameter rates in certain studies underscores the importance of standardized reporting practices for robust assessments in future ocular oncology research.