A deep-learning model for detecting choroidal metastases and predicting primary tumors from ultra-widefield fundus imaging

Abstract

PurposeTo develop and validate a deep-learning model for detecting choroidal metastasis and predicting primary cancer sites using ultra-widefield fundus photography (UWFP).MethodsThis retrospective cohort study utilized 719 UWFP images from 112 patients with choroidal metastasis and 288 normal photos from 288 patients treated at Severance Hospital between 2005 and 2023. A Vision Transformer model, enhanced by transfer learning and image augmentation, was developed and evaluated using AUROC, accuracy, sensitivity, and specificity. Cross-validation, bootstrap sampling, and ablation studies were conducted to ensure robustness and interpretability.ResultsThe model achieved an AUROC of 0.96 for detecting choroidal metastases, significantly outperforming ophthalmologists (AUROC 0.69). Incorporating age and sex information enhanced model performance, yielding AUROCs of 0.87 for lung cancer and 0.96 for breast cancer. Ablation studies confirmed that fundus image features were the primary contributors to classification.ConclusionThe developed deep-learning model shows significant potential not only in detecting choroidal metastases but, more importantly, in predicting their primary cancer origins from UWFP images. This capability could serve as a valuable adjunct in clinical decision-making by guiding more targeted and efficient systemic evaluations, particularly in patients with undiagnosed primary cancers.