Deep learning for early diagnosis of uveal melanoma: a systematic review and meta-analysis

Comprehensive Summary

Uveal melanoma is a rare yet aggressive intraocular cancer with around 6 cases per million people annually. Approximately 50% of patients develop liver metastases with poor prognosis. The researchers aimed to evaluate the usage of deep learning algorithms to detect uveal melanoma. Utilizing PRISMA guidelines, three databases were searched and included studies comparing uveal melanoma patients with healthy individuals or those with choroidal nevi; 5 studies were included consisting of 6,388 patients. The learning model was found to have an overall accuracy of 85% and a positive predictive value of 89%. In addition, the studies analyzed used standard fundus photography and ultra-wide field (UWF) imaging to develop different neural networks. The model was found to have a high sensitivity and good specificity for distinguishing uveal melanoma from benign nevi. The studies were limited in variability and specificity. In conclusion, deep learning demonstrates strong diagnostic performance for detecting uveal melanoma and represents a scalable, noninvasive diagnostic tool.

Outcomes and Implications

This model allows for wider availability of fundus imaging. This analysis can be used to improve early detection and reduce diagnostic issues. Moreover, this imaging is widely available, allowing for a more systematic and widespread diagnostic tool used in areas with less ophthalmology care to standardize care. This one tool is significantly cheaper than excessive referrals and multimodal system which can help optimize resources existing in different healthcare settings.