کنفرانس مهندسی برق ایران

صفحه اصلی / سی و چهارمین کنفرانس بین المللی مهندسی برق

YOLO-Based Deep Learning Framework for Accurate Multi-Class Intraocular Tumor Classification from Ultra-Widefield Fundus Images

نویسندگان :

Helya Haji¹ Sedigheh Dehghani² Mahsa Akhbari³

1- دانشگاه آزاد اسلامی واحد علوم و تحقیقات 2- دانشگاه شهید بهشتی 3- دانشگاه آزاد اسلامی واحد علوم و تحقیقات

کلمات کلیدی :

Deep Learning،YOLO،Intraocular tumor classification،RegNet،EfficientNet

چکیده :

Accurate classification of intraocular tumors is critical for early diagnosis and appropriate treatment planning; however, it remains challenging due to tumor rarity, limited multi-class datasets, pronounced morphological heterogeneity, and overlapping retinal features. In this study, a lightweight YOLO-based deep learning framework is proposed for multi-class classification of five intraocular tumor types using ultra-widefield (UWF) fundus images. The reference dataset study reported a maximum test accuracy of 91.46% using a Vision Transformer (ViT-B) model. Using the baseline YOLO11n-cls architecture, a test accuracy of 94.5% was achieved at an input resolution of 224 × 224 pixels, demonstrating strong performance with a computationally efficient configuration. To further enhance feature representation, the YOLO11n-cls backbone was replaced with advanced convolutional architectures, including EfficientNet-B1 and RegNet-Y-800MF. This resulted in improved test accuracies of 96.0% and 97.5%, respectively, at higher input resolution, with corresponding validation accuracies of 98.2% and 97.2%. These results indicate that attention-enhanced and compound-scaled CNN backbones can significantly improve classification performance while maintaining model efficiency. Overall, the proposed YOLO-based framework demonstrates strong potential for accurate, robust, and computationally efficient automated intraocular tumor diagnosis in clinical settings. In Section I, the clinical importance of accurate intraocular tumor classification is discussed. Previous studies and related work in this domain are reviewed in Section II. The dataset employed in this research and the proposed models are described in Section III. Finally, the obtained classification results are presented and discussed in Section IV.