Abstract
PURPOSE: Accurate assessment and surveillance of retinoblastoma (RB) require more efficient and objective measurements. This study aims to develop an artificial intelligence (AI) system, named RB-Care, for automatic classification and quantitative assessment of RB.
METHODS: A total of 3730 wide-field fundus images were included for the development and validation of 2 models in RB-Care. The first model was trained to automatically classify the images into "normal," "unseeded RB," and "seeded RB." The second model performed quantitative assessment on unseeded RB by detecting and segmenting tumors and optic discs.
RESULTS: The classification model of RB-Care can accurately classify fundus images into 3 categories with an accuracy of 0.9734 and an area under the curve (AUC) of 0.9970. The segmentation model can make precise boundary detection and quantitative measurement on tumors and optic discs, achieving mean Intersection over Union (mIoU) of 0.9670 and Dice similarity coefficient (DSC) of 0.9780 for tumor segmentation, and mIoU of 0.9999 and DSC of 0.9999 for optic disc segmentation, which reaches a comparable level with ophthalmologists.
CONCLUSIONS: The RB-Care achieved excellent performance in both RB classification and segmentation. Consequently, the tumor size and the distance between tumor and optic disc can be quantified, which provides an objective measurement tool for quantitative assessment and surveillance of RB in clinical settings.
TRANSLATIONAL RELEVANCE: Developing a clinically relevant technologies for objective quantitative assessment of RB.