An Integrated Three-Tier Microservice Architecture for Deploying State-of-the-Art Ocular Cancer Detection Models

In this paper, we present our work on the design and deployment of a three-tier scalable microservice-based system architecture for real-time ocular cancer screening based on deep learning models. The platform abstracts the presentation layer (React) from Node data and routing implementation.js, and the specific machine learning inference service that it provides as a FastAPI service. The system enables a strong multistage diagnosis process that now makes use of an EfficientNetV2-B3 transfer learningmodel learned via a two-stage strategy. Trained on a novel re-balanced multi-source dataset of 4,000 retinal images, the model attained an validated accuracy over 97%, with high precision and recall as well as a ROC-AUC of 0.997. This microservice design has been essential for allowing fast, scalable integration of high-performance AI tools into real-world clinical settings– especially in low-resource environments due to robust data persistence, focused patient management, native system traceability and seamless diagnostic request routing.